Augmentation with Dilution? A Large-Scale Empirical Study of Human Contributor Ecosystems After AI Coding Agent Adoption
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: We're looking at the paper: "Augmentation with Dilution? A Large-Scale Empirical Study of Human Contributor Ecosystems After AI Coding Agent Adoption".
Jane: Here's the gist of the paper. AI coding agents are not displacing human contributors, but are systematically reshaping the participation structure of opensource ecosystems <ref:2606.26289#pg2> Using a staggered difference-in-differences design on a dataset of 11,097 GitHub repositories spanning January 2023 to May 2026, we provide causal evidence via the Sun and Abraham estimator <ref:2606.26289#pg5> Our results show that AI agent adoption does not significantly change the absolute number of human contributors (ATT = 0.014, p = 0.224), but significantly reduces human contributor density (ATT = −0.019, p = 0.002) <ref:2606.26289#pg8> The relative participation share of newcomers declines significantly by 3.7 percentage points (AT T = −0.037, p < 0.001), with the effect emerging immediately after adoption and remaining stable throughout the observation window <ref:2606.26289#pg5> Review depth increases significantly by 5.3% (AT T = +0.0168, p < 0.001), indicating that AI agents shift burden from the code production stage to the review stage <ref:2606.26289#pg9> Research Questions and Design The study addresses four research questions regarding the impact of AI coding agent adoption on human contributor participation in open-source projects <ref:2606.26289#pg2>.
Tom: The design employs a staggered difference-in-differences design
12: ,
13: on a dataset of 11,097 GitHub repositories spanning January 2023 to May 2026 <ref:2606.26289#pg5>.
Jane: The primary questions examined are RQ1: How does the adoption of AI coding agents relate to changes in human contributor participation in open-source projects?
Tom: RQ2: Does AI agent adoption reduce the relative participation share of newcomer contributors?
Title and authors: Tom: So we're looking at this paper called "Augmentation with Dilution? A Large-Scale Empirical Study of Human Contributor Ecosystems After AI Coding Agent Adoption." It sounds like they’re asking if these AI coding agents are actually helping or hurting the people who contribute to open source projects.
Jane: Exactly. The title itself, "Augmentation with Dilution," suggests a trade-off, right? It hints that while something is being added—augmentation—something else is shrinking—dilution.
Lu: Right. The authors are looking at this big question using a very specific design to get causal evidence on whether AI agents augment or displace human participation in open source software development > Page two of that work reads: "We present the first largescale empirical study that takes the human contributor ecosystem as its dependent variable, examining how the number, diversity, and behavioral patterns of human contributors change following AI coding agent adoption in open-source projects <ref:2606.26289#pg1,empirical study that takes the human contributor ecosystem as its dependent variable>."
Meng: So they're treating the whole human ecosystem—how many people are contributing and what they do—as the main thing to study, not just the code that’s written by AI > Page two of that work reads: "This paper addresses that gap <ref:2606.26289#pg1>. We present the first largescale empirical study that takes the human contributor ecosystem as its dependent variable, examining how the number, diversity, and behavioral patterns of human contributors change following AI coding agent adoption in open-source projects."
Lalam: I think what's important here is they aren't just looking at code; they are focusing on the people behind that code and their participation patterns > Page two of that work reads: "This paper addresses that gap <ref:2606.26289#pg1>. We present the first largescale empirical study that takes the human contributor ecosystem as its dependent variable, examining how the number, diversity, and behavioral patterns of human contributors change following AI coding agent adoption in open-source projects."
Tom: So to summarize this first part, they're using a staggered difference-in-differences design on eleven thousand repositories from January two thousand twenty-three to May two thousand twenty-six to see if AI adoption changes the absolute number of human contributors or the way they show up <ref:2606.26289#pg1,using a staggered difference-in-differences design on>.
Jane: That's a lot of data spanning over two years, which gives them a good chance to see these effects develop over time > Page one of that work reads: "Using a staggered difference-in-differences design on a dataset of eleven thousand ninety-seven GitHub repositories spanning January two thousand twenty-three to May two thousand twenty-six we provide causal evidence via the Sun and Abraham estimator <ref:2606.26289#pg1,Using a staggered difference-in-differences design on a dataset of 11>."
Lu: And they use the Sun and Abraham estimator for that causal inference part > Page one of that work reads: "Using a staggered difference-in-differences design on a dataset of eleven thousand ninety-seven GitHub repositories spanning January two thousand twenty-three to May two thousand twenty-six we provide causal evidence via the Sun and Abraham estimator <ref:2606.26289#pg1,Using a staggered difference-in-differences design on a dataset of 11>."
Meng: The authors are careful about how they pick their study population. They define a repository as a treatment candidate if it appears in the AIDev dataset with at least one agent-authored pull request and has more than one hundred GitHub stars when they start looking > Page three of that work reads: "We define a repository as a treatment candidate if it satisfies the following quality criteria: (one) it appears in the AIDev dataset with at least one agent-authored pull request, and (two) it has more than one hundred GitHub stars at the time of data collection."
Lalam: That star threshold is important because it excludes those small toy projects, which keeps the study focused on real community activity > Page three of that work reads: "This threshold excludes toy projects and repositories with negligible community activity <ref:2606.26289#pg1>."
Tom: So we’ve got the setup—the data, the time frame, and the specific groups they are tracking—so now we need to see what their actual findings were regarding how AI agents affect these human participants.
The paper's summary: Jane: Okay, so what did this study actually find? They started by looking at four main research questions about how AI coding agent adoption relates to changes in human contributor participation in open source projects > Page two of that work reads: "RQ1: How does the adoption of AI coding agents relate to changes in human contributor participation in open-source projects <ref:2606.26289#pg1>?"
Lu: The main result for the first question is that AI agent adoption does not significantly change the absolute number of human contributors. They found an ATT of zero point zero one four with a p value of point two two four > Page one of that work reads: "Our results show that AI agent adoption does not significantly change the absolute number of human contributors (ATT = zero point zero one four, p = zero point two two four)."
Meng: So they are saying the raw count of human contributors doesn't go up or down significantly just because AI agents are being used > Page one of that work reads: "Our results show that AI agent adoption does not significantly change the absolute number of human contributors (ATT = zero point zero one four, p = zero point two two four)."
Tom: That’s interesting because you might expect some kind of impact on the total headcount if you’re adding a new tool to the workflow > Page one of that work reads: "Our results show that AI agent adoption does not significantly change the absolute number of human contributors (ATT = zero point zero one four, p = zero point two two four)."
Jane: But they move on to something else, and this is where things get more interesting for the ecosystem's structure. They found that AI agent adoption significantly reduces human contributor density by a point minus zero point one nine with a p value of zero point zero two > Page one of that work reads: "but significantly reduces human contributor density (ATT = −zero point zero one nine, p = zero point zero zero two)."
Lu: So they found that the density drops, but they keep the absolute number steady > Page one of that work reads: "but significantly reduces human contributor density (ATT = −zero point zero one nine, p = zero point zero zero two)."
Meng: The density is defined as human contributors divided by pull request count, so this means for every pull request, there are fewer human contributors on average now > Page one of that work reads: "but significantly reduces human contributor density (ATT = −zero point zero one nine, p = zero point zero zero two)."
Tom: That sounds like augmentation with dilution, which is the term they use to describe it later > Page one of that work reads: "This pattern is consistent with augmentation with dilution: human contributors remain present in absolute terms, but their relative share of project contributions declines as AI-generated pull requests accumulate <ref:2606.26289#pg1>."
Jane: And that dilution effect extends to newcomers too. The study found the relative participation share of newcomers declines significantly by three point seven percentage points > Page one of that work reads: "The relative participation share of newcomers declines significantly by three point seven percentage points (ATT = −zero point zero three seven, p < zero point zero zero one)."
Lu: That’s a significant shift in the newcomer ratio, and they say this effect emerges right after the adoption starts and stays stable for the whole observation window > Page five of that work reads: "The relative participation share of newcomers declines significantly by three point seven percentage points (ATT = −zero point zero three seven, p < zero point zero zero one), with the effect emerging immediately after adoption and remaining stable throughout the observation window."
Meng: So they are suggesting that newcomers find fewer entry points when AI agents become common > Page five of that work reads: "The relative participation share of newcomers declines significantly by three point seven percentage points (ATT = −zero point zero three seven, p < zero point zero zero one), with the effect emerging immediately after adoption and remaining stable throughout the observation window."
Tom: It’s a clear picture: fewer total human contributors per pull request, a shrinking share of newcomers, and we've got this pattern of augmentation with dilution to unpack > Page five of that work reads: "The relative participation share of newcomers declines significantly by three point seven percentage points (ATT = −zero point zero three seven, p < zero point zero zero one), with the effect emerging immediately after adoption and remaining stable throughout the observation window."
The paper's improvements: Jane: Now let’s talk about what they suggest we should do next, or what they found in terms of deeper analysis. They looked at how these effects vary across different project characteristics like size and programming language > Page three of that work reads: "RQ4: How do the effects observed in RQ1–RQ3 vary across project characteristics <ref:2606.26289#pg1>."
Lu: They looked at the moderator analysis, which shows these effects aren't uniform across all projects. For example, for Project Size, the effect on newcomer ratio is significant in the low-size group where stars are less than one hundred fifty-six > Page five of that work reads: "For Project Size, the newcomer ratio effect is significant in the low-size group (stars ≤ one hundred fifty-six) with an ATT of −zero point zero seven seven nine (p = zero point zero zero two)."
Meng: So it means AI agents have a stronger shrinking effect on newcomers when a project is small > Page five of that work reads: "For Project Size, the newcomer ratio effect is significant in the low-size group (stars ≤ one hundred fifty-six) with an ATT of −zero point zero seven seven nine (p = zero point zero zero two)."
Tom: And they also looked at programming language, and for Python and TypeScript, the effect on newcomers was consistent across those two languages > Page five of that work reads: "For Programming Language, the effect on newcomer ratio is concentrated in Python and TypeScript, with consistent direction and similar magnitude across the two languages."
Jane: That suggests they should tailor their strategies based on where you are working—like if you're in a small project or using Python or TypeScript > Page five of that work reads: "For Programming Language, the effect on newcomer ratio is concentrated in Python and TypeScript, with consistent direction and similar magnitude across the two languages."
Lu: They also found a similar pattern for project maturity, where the effect on newcomer ratio is only significant in low-maturity groups > Page five of that work reads: "For Project Maturity, the effect on newcomer ratio is significant only in the low-maturity group (AT T = −zero point zero six three six, p."
Meng: So for newcomers who are joining small or young projects, their opportunity shrinks faster > Page five of that work reads: "For Project Maturity, the effect on newcomer ratio is significant only in the low-maturity group (AT T = −zero point zero six three six, p."
Tom: So the paper points toward a tailored approach—you can't treat every project the same way when it comes to newcomer retention > Page five of that work reads: "For Project Maturity, the effect on newcomer ratio is significant only in the low-maturity group (AT T = −zero point zero six three six, p."
Jane: And they also touched on review depth. They found that review depth increases significantly by five point three percent because AI agents shift the burden from writing code production to the review stage > Page nine of that work reads: "Review depth increases significantly by five point three percent (AT T = +zero point zero one six eight, p < zero point zero zero one)."
Lu: So they’re showing that reviewers are spending more time reviewing, which is a direct consequence of AI agents taking over some of the initial coding tasks > Page nine of that work reads: "Review depth increases significantly by five point three percent (AT T = +zero point zero one six eight, p < zero point zero zero one)."
Meng: That means the burden shifts from writing code to assessing it > Page nine of that work reads: "Review depth increases significantly by five point three percent (AT T = +zero point zero one six eight, p < zero point zero zero one)."
Tom: So we’ve got the dilution effect on newcomers, the change in density, and now we're seeing more review effort shifting to maintainer work > Page nine of that work reads: "Review depth increases significantly by five point three percent (AT T = +zero point zero one six eight, p < zero point zero zero one)."
Conclusion: Jane: So, to wrap up the findings on this paper titled "Augmentation with Dilution? A Large-Scale Empirical Study of Human Contributor Ecosystems After AI Coding Agent Adoption," the main message is that AI agents aren't replacing humans in terms of absolute numbers.
Lu: That’s right, they don't significantly change the absolute number of human contributors, but they are definitely associated with declining human contributor density and a shrinking relative participation share for newcomers > Page one of that work reads: "Our results show that AI agent adoption does not significantly change the absolute number of human contributors (ATT = zero point zero one four, p = zero point two two four), but significantly reduces human contributor density (ATT = −zero point zero one nine, p = zero point zero zero two)."
Tom: And that pattern they call augmentation with dilution—humans are still around in absolute terms, but their relative share is shrinking as AI-generated pull requests pile up > Page one of that work reads: "This pattern is consistent with augmentation with dilution: human contributors remain present in absolute terms, but their relative share of project contributions declines as AI-generated pull requests accumulate <ref:2606.26289#pg1>."
Jane: And they point out that the effect on newcomers is strongest in smaller projects and for Python and TypeScript languages > Page five of that work reads: "For Project Size, the newcomer ratio effect is significant in the low-size group (stars ≤ one hundred fifty-six) with an ATT of −zero point zero seven seven nine (p = zero point zero zero two)."
Lu: And they also found that review depth increases by five point three percent because AI agents shift the burden from code production to the review stage > Page nine of that work reads: "Review depth increases significantly by five point three percent (AT T = +zero point zero one six eight, p < zero point zero zero one)."
Meng: From my side, I’m concerned about what this means for us in production; we need to think about how to manage that increased review load since the burden is shifting away from just writing new code > Page nine of that work reads: "Review depth increases significantly by five point three percent (AT T = +zero point zero one six eight, p < zero point zero zero one)."
Tom: So we’ve seen a clear pattern across the whole dataset for the paper "Augmentation with Dilution? A Large-Scale Empirical Study of Human Contributor Ecosystems After AI Coding Agent Adoption," and it suggests that the way human participation works is evolving, not necessarily disappearing > Page one of that work reads: "This paper presents the first large-scale empirical study that takes the human contributor ecosystem as its dependent variable, examining how the number, diversity, and behavioral patterns of human contributors change following AI coding agent adoption in open-source projects <ref:2606.26289#pg1,This paper presents the first large-scale empirical study that takes the>."
Jane: It really shows that while AI agents are a powerful tool for generating code, they’re reshaping how we measure success in an ecosystem > Page one of that work reads: "This paper presents the first large-scale empirical study that takes the human contributor ecosystem as its dependent variable, examining how the number, diversity, and behavioral patterns of human contributors change following AI coding agent adoption in open-source projects <ref:2606.26289#pg1,This paper presents the first large-scale empirical study that takes the>."
Lu: I think this is important because it gives us concrete evidence about the dilution effect on newcomers and where to focus our efforts > Page one of that work reads: "This paper presents the first large-scale empirical study that takes the human contributor ecosystem as its dependent variable, examining how the number, diversity, and behavioral patterns of human contributors change following AI coding agent adoption in open-source projects <ref:2606.26289#pg1,This paper presents the first large-scale empirical study that takes the>."
Meng: So we need to keep an eye on how this trend plays out as AI becomes more common in our workflows > Page one of that work reads: "This paper presents the first large-scale empirical study that takes the human contributor ecosystem as its dependent variable, examining how the number, diversity, and behavioral patterns of human contributors change following AI coding agent adoption in open-source projects <ref:2606.26289#pg1,This paper presents the first large-scale empirical study that takes the>."
Tom: We’ll take this insight about augmentation with dilution from this paper and see what we can do next for our own workflows.
Weixing Zhang, Bowen Jiang, Anne Koziolek
Karlsruhe Institute of Technology
cs.SE
Submitted: 2026-06-24
Updated: 2026-11-05
Comments: v2: Substantially revised. Outcome variables are now computed from complete pull request histories retrieved through the GitHub API instead of GHArchive, whose coverage of pull request events declines from mid-2025. The decrease in the newcomer share reported in v1 is not supported in this version, and the review-depth analysis has been removed
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 86/100
The gist: AI coding agents are not displacing human contributors, but are systematically reshaping the participation structure of opensource ecosystems <ref:2606.26289#pg2> Using a staggered…
Key concepts
- Human Contributor Density
- This measures how many human developers are contributing relative to the number of pull requests in a project. A decrease in this density means that as AI generates more code, the human contribution becomes less frequent compared to the total output.
- Newcomer Relative Share
- This tracks what proportion of contributions comes from new contributors. The study found this share drops significantly after AI adoption, indicating that newcomers are less likely to participate relative to established developers.
- Augmentation with Dilution
- This is the core finding: AI agents augment (help) the overall project output but dilute (reduce the relative importance of) human input. Human contributors are still present in total numbers, but their proportional share shrinks as AI takes over tasks.
Terminology
Summary
AI coding agents are not displacing human contributors, but are systematically reshaping the participation structure of opensource ecosystems <ref:2606. p. 0 Together, these findings present a pattern of augmentation with
dilution: AI agents are not displacing human contributors, but
are systematically reshaping the participation structure of opensource ecosystems. Index Terms—AI Coding Agents, Open-Source Software, Human Contributors, Difference-in-Differences, Newcomer Participation. This paper presents the
first large-scale empirical study that takes the human contributor
ecosystem as its dependent variable, examining how the number, composition, and behavior of human participants change
following AI coding agent adoption in open-source projects. Using a staggered difference-in-differences design on a dataset
of 11,097 GitHub repositories spanning January 2023 to May
2026, we provide causal evidence via the Sun and Abraham
estimator. Using a staggered difference-in-differences design on a dataset
of 11,097 GitHub repositories spanning January 2023 to May
2026, we provide causal evidence via the Sun and Abraham
estimator. Our results show that AI agent adoption does not
significantly change the absolute number of human contributors
(AT T = 0. Our results show that AI agent adoption does not
significantly change the absolute number of human contributors
(AT T = 0. 014, p = 0. Our results show that AI agent adoption does not
significantly change the absolute number of human contributors
(AT T = 0. 014, p = 0. 014, p = 0. 224), but significantly reduces human
contributor density (AT T = −0. 224), but significantly reduces human
contributor density (AT T = −0. 019, p = 0. 224), but significantly reduces human
contributor density (AT T = −0. 019, p = 0. These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606. These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606. These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606. 002), indicating
that the relative share of human participation declines as AIgenerated pull requests accumulate. The relative participation
share of newcomers declines significantly by 3. The relative participation
share of newcomers declines significantly by 3. 7 percentage
points (AT T = −0. 7 percentage
points (AT T = −0. 037, p < 0. 037, p < 0. 001), with the effect emerging
immediately after adoption and remaining stable throughout
the observation window. 001), with the effect emerging
immediately after adoption and remaining stable throughout
the observation window. Review depth increases significantly by
5. 3% (AT T = +0. 0168, p < 0. 0168, p < 0. 001), indicating that AI agents
shift burden from the code production stage to the review stage.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Research Questions and Design
The study addresses four research questions regarding the impact of AI coding agent adoption on human contributor participation in open-source projects <ref:2606.26289#pg2>. The design employs a staggered difference-in-differences design [12], [13] on a dataset of 11,097 GitHub repositories spanning January 2023 to May 2026 <ref:2606. p. 0 This paper presents the
first large-scale empirical study that takes the human contributor
ecosystem as its dependent variable, examining how the number, composition, and behavior of human participants change
following AI coding agent adoption in open-source projects. Using a staggered difference-in-differences design on a dataset
of 11,097 GitHub repositories spanning January 2023 to May
2026, we provide causal evidence via the Sun and Abraham
estimator. To the best of our knowledge, this is
the first study to treat the human contributor ecosystem, rather
than AI-generated artifacts, as the primary lens through which
to evaluate the impact of AI coding agents on open-source
software development. RQ1: How does the adoption of AI coding agents relate
to changes in human contributor participation in open-source
projects? RQ1: How does the adoption of AI coding agents relate
to changes in human contributor participation in open-source
projects? RQ2: Does AI agent adoption reduce the relative participation share of newcomer contributors? RQ2: Does AI agent adoption reduce the relative participation share of newcomer contributors? RQ3: Does AI coding agent adoption change the depth of
review provided per pull request? This RQ examines whether
reviewers invest greater per-review effort as agent-generated
pull requests accumulate. RQ4: How do the effects observed in RQ1–RQ3 vary across
project characteristics?Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Outcome Variables and Causal Strategy
Two complementary outcome variables are defined for RQ1: the log-transformed monthly human contributor count, defined as log(human contributors + 1), and human contributor density (hc density), defined as human contributors/pr count <ref:2606.26289#pg6>. For RQ2, measures include the log-transformed monthly newcomer count and newcomer ratio, which is newcomers/human contributors <ref:2606. p. 4 For RQ2, we define two complementary outcome variables. The first measure is the log-transformed monthly
newcomer count, defined as log(newcomers + 1), where a
newcomer is a user who submits a pull request to a given
repository in a given month and has no prior pull request
activity in that repository between January 2020 and the
month in question. Both measures are estimated using the same Sun &
Abraham estimator applied in RQ1; where the two measures
yield divergent parallel trends diagnostics, the measure satisfying the parallel trends assumption is treated as the primary
causal evidence. Causal Inference Strategy
We employ a staggered difference-in-differences (DiD) design to estimate the causal effect of AI agent adoption on
human contributor participation. This estimator yields
more robust estimates than TWFE when treatment effects are
heterogeneous across cohorts. The model includes repository fixed effects and month fixed
effects, absorbing time-invariant repository-level heterogeneity and common time trends shared across all repositories,
respectively.Where the paper says it
Where the paper says it
Where the paper says it
Key Findings on Participation Dynamics
The main results indicate that AI agent adoption does not significantly reduce the absolute number of human contributors (ATT = 0.014, p = 0.224) <ref:2606. p. 0 These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606. 014, p = 0. 224), but significantly decreases
human contributor density by approximately 1. 9 percentage points in human
contributor density. For RQ1-A, the aggregated ATT = 0. 019, p = 0. This pattern
is consistent with augmentation with dilution: human
contributors remain present in absolute terms, but their
relative share of project contributions declines as AIgenerated pull requests accumulate. : The event study coefficients in
Figure 3 show that the reduction in newcomer ratio emerges
from period +1 onward and remains stable throughout the
post-adoption observation window, with no sign of attenuation. Answer to RQ2: AI agent adoption is associated with a
significant and sustained reduction in newcomer relative
share of approximately 3. 9 percentage points in human
contributor density. For RQ1-A, the aggregated ATT = 0. 6% from the pre-adoption baseline. The effect emerges
immediately after adoption and remains stable throughout
the observation window.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Moderator Analysis and Heterogeneity
Moderator analysis reveals that these effects vary systematically with project size, programming language, and project maturity <ref:2606.26289#pg9>. For Project Size, the newcomer ratio effect is significant in the low-size group (stars ≤ 156) with an ATT of −0. p. 7 Project Size. For newcomer ratio, the low-size group
(stars ≤ 156) shows a significant negative effect (AT T =
−0. For newcomer ratio, the low-size group
(stars ≤ 156) shows a significant negative effect (AT T =
−0. 0779, p = 0. 0540, p < 0. 001), with consistent
direction and similar magnitude across the two languages. The maturity subgroups exhibit a pattern
distinct from the size subgroups. The effect on newcomer
ratio is significant only in the low-maturity group (AT T =
−0.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Review Effort Shifts
RQ3 examines whether review effort per pull request changes following AI agent adoption <ref:2606.26289#pg9>. The regression results show that review depth (review depth) increases by 0. p. 6 RQ3: Maintainer Review Effort After AI Agent Adoption
RQ3 examines whether review effort per pull request
changes following AI agent adoption. The regression results
show that review depth (review depth) increases by 0. The regression results
show that review depth (review depth) increases by 0. 0168 on
average after AI agent adoption (SE = 0. Post-treatment coefficients remain consistently positive, with the effect emerging immediately after adoption and
persisting without attenuation or reversal over time. This result indicates that after AI agent adoption, reviewers
provide more comments on average per pull request. This result indicates that after AI agent adoption, reviewers
provide more comments on average per pull request. Combined with the findings from RQ1 (declining human contributor density) and RQ2 (declining newcomer ratio), the three
results jointly present a consistent picture: AI agent adoption
accelerates code production, while the relative share of human
participants in project activity is declining, and the remaining
core members invest more effort in the review process.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Discussion and Implications
The causal estimates for RQ2 show a sustained decline in newcomer relative share following AI agent adoption, but the mechanism behind this finding warrants further discussion <ref:2606.26289#pg10>. Two non-mutually exclusive explanations exist for the reduction in newcomer ratio: a dilution effect, where the denominator increases mechanically reducing the newcomer share, and a displacement effect, where AI agents handle well-scoped tasks historically served as entry points <ref:2606. p. 7 The first is a dilution effect: as
AI agents generate pull requests at scale, the denominator of
the ratio increases, mechanically reducing the newcomer share
even if the absolute number of newcomers remains unchanged. The second is a displacement effect: AI agents tend to handle
well-scoped, moderately complex tasks, precisely the task
types that have historically served as entry points through Future work examining the task-type distributions of AI
agent contributions relative to first-time human contributions
could provide a more direct test of the displacement hypothesis. The concentration of the newcomer dilution effect in Python
and TypeScript warrants further investigation; one possible
explanation is that AI agent adoption density is higher in
these ecosystems, but this conjecture requires language-level
adoption rate data to verify. The concentration of the newcomer dilution effect in Python
and TypeScript warrants further investigation; one possible
explanation is that AI agent adoption density is higher in
these ecosystems, but this conjecture requires language-level
adoption rate data to verify. Regarding RQ3, a potential
concern is whether the increase in review depth reflects a
shift in maintainer behavior or is driven by a small number of
repositories with disproportionately high review activity. 002), while neither outcome shows
a significant effect in the low-maturity group. Answer to RQ4: The effect of AI agent adoption on newcomer relative share is stronger in smaller and younger
projects, and is most concentrated in the Python and
TypeScript ecosystems. Using a staggered difference-indifferences design on 11,097 GitHub repositories, we find that
AI agent adoption does not significantly reduce the absolute
number of human contributors, but is associated with declining
human contributor density, a shrinking relative participation
share of newcomers, and increased per-review depth among
observed reviewers. These three findings jointly present a
pattern of augmentation with dilution: AI agents reshape
the participation structure of open-source ecosystems without
displacing human contributors in absolute terms.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Threats to Validity
The paper notes several threats to validity, including using pull request submissions as a proxy for human contributor activity, which may lead to an underestimation of human contributor counts <ref:2606.26289#pg10>. The bot filtering relies on string matching against contributor login names, and residual threats cannot be eliminated without manual inspection <ref:2606. p. 9 Our bot filtering relies on string matching against contributor login names, specifically excluding accounts containing
“[bot]” or “bot” and three commonly used automation accounts. Bot accounts that do not conform to these naming
conventions may be incorrectly counted as human contributors; this residual threat cannot be eliminated without manual
inspection of all accounts. However, for contributors who had contribution history in a
repository prior to January 2020, our data may still incorrectly
classify them as newcomers, potentially leading to a slight
overestimation of newcomer counts. The primary programming language of control repositories was obtained via the GitHub REST API rather than
from GHArchive, as GHArchive event data does not include repository-level language metadata. This operationalization is subject to measurement error: a repository may have begun using AI coding
agents prior to the appearance of its first AIDev-recorded pull
request, either because earlier agent-authored contributions
were not captured by AIDev’s detection heuristics or because
the repository adopted an agent that generated contributions
outside the scope of the dataset. This measurement error
implies that our treatment onset is a lower bound on true
adoption timing, and that the pre-treatment period for some
repositories may in fact contain post-adoption observations. c) External Validity: Our treatment group consists of
repositories with more than 100 GitHub stars, which excludes
smaller projects with lower visibility. The findings of this study
may not generalize to repositories below this threshold, as the
dynamics of AI agent adoption and its effects on contributor
ecosystems may differ in smaller or less active projects.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
This paper provides the first causal evidence on the effects of AI coding agent adoption on human contributor ecosystems in open-source projects <ref:2606.26289#pg10>. Using a staggered difference-indifferences design on 11,097 GitHub repositories, we find that AI agent adoption does not significantly reduce the absolute number of human contributors, but is associated with declining human contributor density, a shrinking relative participation share of newcomers, and increased per-review depth among observed reviewers. p. 9 CONCLUSION
This paper provides the first causal evidence on the effects
of AI coding agent adoption on human contributor ecosystems in open-source projects. Using a staggered difference-indifferences design on 11,097 GitHub repositories, we find that
AI agent adoption does not significantly reduce the absolute
number of human contributors, but is associated with declining
human contributor density, a shrinking relative participation
share of newcomers, and increased per-review depth among
observed reviewers. Using a staggered difference-indifferences design on 11,097 GitHub repositories, we find that
AI agent adoption does not significantly reduce the absolute
number of human contributors, but is associated with declining
human contributor density, a shrinking relative participation
share of newcomers, and increased per-review depth among
observed reviewers. These three findings jointly present a
pattern of augmentation with dilution: AI agents reshape
the participation structure of open-source ecosystems without
displacing human contributors in absolute terms. 001), indicating that AI agents
shift burden from the code production stage to the review stage. Moderator analysis reveals that these effects vary systematically
with project size, programming language, and project maturity. Moderator
analysis reveals that these effects vary with project size,
programming language, and project maturity. As AI agents
see widespread adoption across open-source ecosystems, the
evolution of human participation structures warrants continued
attention from the research community. As AI agents
see widespread adoption across open-source ecosystems, the
evolution of human participation structures warrants continued
attention from the research community. Future work may
examine the newcomer displacement hypothesis at the tasktype level, and extend the research lens to enterprise software
development environments.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
References
[1] R. Robbes, T. Matricon, T. Degueule, A. Hora, and S. Zacchiroli, “Agentic much? p. 0 There
is already suggestive evidence: over 67. 5% of AI-co-authored
pull requests originate from contributors without prior code
ownership [8], indicating that a disproportionate share of
agent-generated contributions come from contributors with
no prior ownership history in the repository; and METR’s
longitudinal study of open-source developers noted that as
agentic tool adoption increased throughout 2025, “recruitment
and retention of developers has become more difficult” [10]. These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606. These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606.Where the paper says it
Where the paper says it
Where the paper says it
[2] I. Steinmacher, M. p. 1 b) Human Contributor Ecosystems in Open-Source Software: The long-term sustainability of open-source projects
is closely tied to the health of their contributor ecosystems. Steinmacher et al. Steinmacher et al. [2] identified, through a systematic literature review, the barriers newcomers face when joining opensource projects, establishing the importance of a sustained
newcomer influx to project survival.Where the paper says it
Where the paper says it
[3] C. Miller, D. G. Widder, C. Kastner, and B. Vasilescu, “Why do people ¨ give up flossing? p. 1 [3] examined
the mechanisms underlying core contributor disengagement,
revealing the relationship between maintainer burden and
contributor attrition. Vasilescu et al. edu
Abstract—AI coding agents are penetrating open-source software development at an unprecedented pace, yet existing research
predominantly treats human contributors as a static backdrop
rather than as the subject of inquiry. This paper presents the
first large-scale empirical study that takes the human contributor
ecosystem as its dependent variable, examining how the number, composition, and behavior of human participants change
following AI coding agent adoption in open-source projects. These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606.Where the paper says it
Where the paper says it
Where the paper says it
[4] B. Vasilescu, D. p. 1 [3] examined
the mechanisms underlying core contributor disengagement,
revealing the relationship between maintainer burden and
contributor attrition. Vasilescu et al. We present the first largescale empirical study that takes the human contributor ecosystem as its dependent variable, examining how the number, diversity, and behavioral patterns of human contributors change
following AI coding agent adoption in open-source projects. Using a staggered difference-in-differences design [12], [13]
on a dataset of 11,097 GitHub repositories spanning January
2023 to May 2026, we provide causal evidence on whether
AI agent adoption augments or displaces human participation,
whether newcomers’ relative participation share is shrinking,
how review effort per pull request shifts as AI-generated
contributions accumulate, and how these effects vary across
project characteristics.Where the paper says it
Where the paper says it
[5] D. Ogenrwot and J. p. 1 The
first characterizes the structural properties of agent-generated
contributions: Robbes et al. [1] quantified the adoption scale
of coding agents on GitHub; Ogenrwot and Businge [5],
Rahman and Shihab [6], and Liu et al. [1] quantified the adoption scale
of coding agents on GitHub; Ogenrwot and Businge [5],
Rahman and Shihab [6], and Liu et al. [7] characterized agentgenerated pull requests at scale from the perspectives of code
structure, survival rate, and technical debt, respectively; Ehsani
et al. These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606.Where the paper says it
Where the paper says it
Where the paper says it
[6] M. Rahman and E. p. 1 The
first characterizes the structural properties of agent-generated
contributions: Robbes et al. [1] quantified the adoption scale
of coding agents on GitHub; Ogenrwot and Businge [5],
Rahman and Shihab [6], and Liu et al. The
first characterizes the structural properties of agent-generated
contributions: Robbes et al. [1] quantified the adoption scale
of coding agents on GitHub; Ogenrwot and Businge [5],
Rahman and Shihab [6], and Liu et al. These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606. These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[7] Y. Liu, R. Widyasari, Y. Zhao, I. C. Irsan, J. Chen, and D. Lo, “Debt behind the ai boom: A large-scale empirical study of ai-generated code in the wild,” arXiv preprint arXiv:2603. p. 0 5% of AI-co-authored
pull requests originate from contributors without prior code
ownership [8], indicating that a disproportionate share of
agent-generated contributions come from contributors with
no prior ownership history in the repository; and METR’s
longitudinal study of open-source developers noted that as
agentic tool adoption increased throughout 2025, “recruitment
and retention of developers has become more difficult” [10]. At
the same time, the large-scale influx of agent-generated pull
requests restructures the review workload facing core maintainers [11]: as the proportion of AI-generated contributions
in the PR pool grows, maintainers may need to exercise greater
judgment on each review to assess correctness, integration risk,
and contextual fit, burdens that do not diminish simply because
a large share of PRs are merged without explicit review [8]. These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606.Where the paper says it
Where the paper says it
[8] H Gao, P. p. 1 [15] analyzed the patterns and causes of failed agentgenerated pull requests. The second examines the effects of
agent adoption on project-level outcomes: Gao et al. We present the first largescale empirical study that takes the human contributor ecosystem as its dependent variable, examining how the number, diversity, and behavioral patterns of human contributors change
following AI coding agent adoption in open-source projects. Using a staggered difference-in-differences design [12], [13]
on a dataset of 11,097 GitHub repositories spanning January
2023 to May 2026, we provide causal evidence on whether
AI agent adoption augments or displaces human participation,
whether newcomers’ relative participation share is shrinking,
how review effort per pull request shifts as AI-generated
contributions accumulate, and how these effects vary across
project characteristics. These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606.Where the paper says it
Where the paper says it
Where the paper says it
[9] S. Agarwal, H He, and B Vasilescu, “AI IDEs or autonomous agents? p. 1 [3] examined
the mechanisms underlying core contributor disengagement,
revealing the relationship between maintainer burden and
contributor attrition. Vasilescu et al. Using a staggered difference-in-differences design [12], [13]
on a dataset of 11,097 GitHub repositories spanning January
2023 to May 2026, we provide causal evidence on whether
AI agent adoption augments or displaces human participation,
whether newcomers’ relative participation share is shrinking,
how review effort per pull request shifts as AI-generated
contributions accumulate, and how these effects vary across
project characteristics. To the best of our knowledge, this is
the first study to treat the human contributor ecosystem, rather
than AI-generated artifacts, as the primary lens through which
to evaluate the impact of AI coding agents on open-source
software development.Where the paper says it
Where the paper says it
[10] J. Becker, N. Rush, T. Cunningham, D Rein, and K Mahamud, “We are changing our developer productivity experiment design,” https://metr. p. 0 There
is already suggestive evidence: over 67. 5% of AI-co-authored
pull requests originate from contributors without prior code
ownership [8], indicating that a disproportionate share of
agent-generated contributions come from contributors with
no prior ownership history in the repository; and METR’s
longitudinal study of open-source developers noted that as
agentic tool adoption increased throughout 2025, “recruitment
and retention of developers has become more difficult” [10]. This paper presents the
first large-scale empirical study that takes the human contributor
ecosystem as its dependent variable, examining how the number, composition, and behavior of human participants change
following AI coding agent adoption in open-source projects. Using a staggered difference-in-differences design on a dataset
of 11,097 GitHub repositories spanning January 2023 to May
2026, we provide causal evidence via the Sun and Abraham
estimator.Where the paper says it
Where the paper says it
[11] G. Gousios, A Zaidman, M.-A Storey, and A Van Deursen, “Work practices and challenges in pull-based development: The integrator’s perspective,” in 2015 IEEE/ACM 37th IEEE International Conference on Software Engineering <ref:2606. p. 1 The replication package, including all analysis scripts and
data, is available at [14]. RELATED WORK
a) AI Coding Agents in Open-Source Development: Empirical research on AI coding agents in open-source software
has grown rapidly, forming two main lines of inquiry.Where the paper says it
[12] B Callaway and P. H Sant’Anna, “Difference-in-differences with multiple time periods,” Journal of econometrics, vol 225, no 2, pp 200–230, 2021 <ref:2606. p. 3 While
three covariates marginally exceed the conventional threshold of 0. 10, the difference-in-differences estimator controls
for time-invariant repository characteristics through repository
fixed effects, which absorbs much of the remaining imbalance.Where the paper says it
[13] C A Furia, R Torkar, and R Feldt, “Towards causal analysis of empirical software engineering data: The impact of programming languages on coding competitions,” ACM Transactions on Software Engineering and Methodology, vol 33, no 1, pp 1–35, 2023 <ref:2606. p. 1 Using a staggered difference-in-differences design [12], [13]
on a dataset of 11,097 GitHub repositories spanning January
2023 to May 2026, we provide causal evidence on whether
AI agent adoption augments or displaces human participation,
whether newcomers’ relative participation share is shrinking,
how review effort per pull request shifts as AI-generated
contributions accumulate, and how these effects vary across
project characteristics. To the best of our knowledge, this is
the first study to treat the human contributor ecosystem, rather
than AI-generated artifacts, as the primary lens through which
to evaluate the impact of AI coding agents on open-source
software development.Where the paper says it
[14] A Author, “Replication package for ‘Augmentation with Dilution: A Large-Scale Empirical Study of Human Contributor Ecosystems After AI Coding Agent Adoption’,” 2026 <ref:2606. p. 0 Augmentation with Dilution: A Large-Scale
Empirical Study of Human Contributor Ecosystems
After AI Coding Agent Adoption
1
st Weixing Zhang
Karlsruhe Institute of Technology
Karlsruhe, Germany
weixing. edu
2
nd Bowen Jiang
Karlsruhe Institute of Technology
Karlsruhe, Germany
bowen.Where the paper says it
[15] R Ehsani, S Pathak, S Rawal, A A Mujahid, M M Imran, and P Chatterjee, “Where do ai coding agents fail? p. 1 [1] quantified the adoption scale
of coding agents on GitHub; Ogenrwot and Businge [5],
Rahman and Shihab [6], and Liu et al. [7] characterized agentgenerated pull requests at scale from the perspectives of code
structure, survival rate, and technical debt, respectively; Ehsani
et al. These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606. These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606.Where the paper says it
Where the paper says it
Where the paper says it
[16] M Watanabe, H Li, Y Kashiwa, B Reid, H Iida, and A E Hassan, “On the use of agentic coding: An empirical study of pull requests on github,” ACM Transactions on Software Engineering and Methodology <ref:2606. p. 0 Yet existing empirical research on AI coding agents shares
a systematic blind spot: without exception, it treats human
contributors as a static backdrop rather than as the subject of
inquiry. Prior work has characterized the structural properties
of AI-generated pull requests [5], measured the code quality
and survival rates of AI-authored contributions [6], quantified
the technical debt they introduce [7], studied how AI-generated
code is reviewed, or more often, not reviewed [8], and provided the first causal evidence that agent adoption degrades
software quality at the project level [9].Where the paper says it
[17] M Monperrus, “The end of code review: Coding agents supersede human inspection,” arXiv preprint arXiv:2606. p. 0 These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606. These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606.Where the paper says it
Where the paper says it
[18] M Valiev, B Vasilescu, and J Herbsleb, “Ecosystem-level determinants of sustained activity in open-source projects: A case study of the pypi ecosystem,” in Proceedings of the 2018 26th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering <ref:2606. p. 0 edu
Abstract—AI coding agents are penetrating open-source software development at an unprecedented pace, yet existing research
predominantly treats human contributors as a static backdrop
rather than as the subject of inquiry. This paper presents the
first large-scale empirical study that takes the human contributor
ecosystem as its dependent variable, examining how the number, composition, and behavior of human participants change
following AI coding agent adoption in open-source projects.Where the paper says it
[19] H Li, H Zhang, and A E Hassan, “The rise of AI teammates in software engineering (SE) 3. p. 0 Augmentation with Dilution: A Large-Scale
Empirical Study of Human Contributor Ecosystems
After AI Coding Agent Adoption
1
st Weixing Zhang
Karlsruhe Institute of Technology
Karlsruhe, Germany
weixing. edu
2
nd Bowen Jiang
Karlsruhe Institute of Technology
Karlsruhe, Germany
bowen. Together, these findings present a pattern of augmentation with
dilution: AI agents are not displacing human contributors, but
are systematically reshaping the participation structure of opensource ecosystems. Index Terms—AI Coding Agents, Open-Source Software, Human Contributors, Difference-in-Differences, Newcomer Participation.Where the paper says it
Where the paper says it
[20] H He, C Miller, S Agarwal, C Kastner, and B Vasilescu, “Speed at ¨ the cost of quality: How Cursor AI increases short-term velocity and long-term complexity in open-source projects,” in Proceedings of the 23rd International Conference on Mining Software Repositories (MSR), 2026 <ref:2606. p. 1 Monperrus [17] argues on
the basis of capability evidence that human code review can
already be replaced by AI agents; the empirical results of RQ3
in the present study point in the opposite direction: review
depth increases significantly following AI agent adoption,
indicating that human review burden grows as agent-generated
contributions accumulate. b) Human Contributor Ecosystems in Open-Source Software: The long-term sustainability of open-source projects
is closely tied to the health of their contributor ecosystems.Where the paper says it
[21] P R Rosenbaum and D B Rubin, “The central role of the propensity score in observational studies for causal effects,” Biometrika, vol 70, no 1, pp 41–55, 1983 <ref:2606. p. 9 If this condition
holds, our findings are unlikely to overstate the effects of AI
agent adoption on human contributor ecosystems. b) Internal Validity: The propensity score matching procedure achieved acceptable but imperfect covariate balance,
which bears directly on the comparability of treatment and
control repositories.Where the paper says it
[22] P C Austin, “An introduction to propensity score methods for reducing the effects of confounding in observational studies,” Multivariate behavioral research, vol 46, no 3, pp 399–424, 2011 <ref:2606.26289#pg10>.
[23] J M Wooldridge, “Two-way fixed effects, the two-way mundlak regression, and difference-in-differences estimators,” Empirical Economics, vol 69, no 5, pp 2545–2587, 2025 <ref:2606. p. 4 Because different repositories
adopted AI agents in different months, the data exhibit staggered treatment timing. Under treatment effect heterogeneity
across cohorts, the traditional two-way fixed effects (TWFE)
estimator can produce biased estimates, as early-adopting
repositories are incorrectly used as controls for late-adopting
ones [23].Where the paper says it
[24] L Sun and S Abraham, “Estimating dynamic treatment effects in event studies with heterogeneous treatment effects,” Journal of econometrics, vol 225, no 2, pp 175–199, 2021 <ref:2606. p. 4 We therefore adopt the interaction-weighted estimator proposed by Sun and Abraham [24], which decomposes the
aggregate average treatment effect on the treated (ATT) into a
weighted average of cohort-specific average treatment effects
(CATTs) at each relative time period. This estimator yields
more robust estimates than TWFE when treatment effects are
heterogeneous across cohorts.Where the paper says it
[25] D S D Minh, H T Kiet, N L P Quy, P P Hoa, T C Nguyen, N D H Duong, and T B Tran, “Early-stage prediction of review effort in ai-generated pull requests,” arXiv preprint arXiv:2601. p. 1 When approximately 80% of AI-generated PRs are merged
without explicit review [8], the review effort that does occur
may become more intensive. This RQ examines whether
reviewers invest greater per-review effort as agent-generated
pull requests accumulate. These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606.Where the paper says it
Where the paper says it
[26] J Tsay, L Dabbish, and J Herbsleb, “Influence of social and technical factors for evaluating contribution in github,” in Proceedings of the 36th international conference on Software engineering <ref:2606. p. 0 Yet existing empirical research on AI coding agents shares
a systematic blind spot: without exception, it treats human
contributors as a static backdrop rather than as the subject of
inquiry. Prior work has characterized the structural properties
of AI-generated pull requests [5], measured the code quality
and survival rates of AI-authored contributions [6], quantified
the technical debt they introduce [7], studied how AI-generated
code is reviewed, or more often, not reviewed [8], and provided the first causal evidence that agent adoption degrades
software quality at the project level [9].Where the paper says it
[27] K Chowdhury, D Banik, K Ferdous, and S I Shamim, “From industry claims to empirical reality: An empirical study of code review agents in pull requests,” arXiv preprint arXiv:2604. p. 0 Yet existing empirical research on AI coding agents shares
a systematic blind spot: without exception, it treats human
contributors as a static backdrop rather than as the subject of
inquiry. Prior work has characterized the structural properties
of AI-generated pull requests [5], measured the code quality
and survival rates of AI-authored contributions [6], quantified
the technical debt they introduce [7], studied how AI-generated
code is reviewed, or more often, not reviewed [8], and provided the first causal evidence that agent adoption degrades
software quality at the project level [9]. These observations, drawn from studies with different foci
and datasets, point collectively toward a consequential phenomenon that no study has yet examined directly. arXiv:2606.Where the paper says it
Where the paper says it
[28] G Gousios, M Pinzger, and A v Deursen, “An exploratory study of the pull-based software development model,” in Proceedings of the 36th international conference on software engineering <ref:2606. p. 1 Using a staggered difference-in-differences design [12], [13]
on a dataset of 11,097 GitHub repositories spanning January
2023 to May 2026, we provide causal evidence on whether
AI agent adoption augments or displaces human participation,
whether newcomers’ relative participation share is shrinking,
how review effort per pull request shifts as AI-generated
contributions accumulate, and how these effects vary across
project characteristics. To the best of our knowledge, this is
the first study to treat the human contributor ecosystem, rather
than AI-generated artifacts, as the primary lens through which
to evaluate the impact of AI coding agents on open-source
software development.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improve development pipelines by implementing mechanisms that manage AI-generated pull request volume growth to counteract dilution effects on human participation, as this
mechanically reduc[es] the denominator
of human contributor density. -
Enhance newcomer onboarding by developing targeted strategies for entry-level tasks, as the paper suggests agents
systematically occupy these tasks,
potentially shrinking contribution opportunities for newcomers. -
Optimize code review workflows by shifting maintainer focus from basic code production to higher-judgment assessment, as
AI agents shift burden from the code production stage to the review stage
andreview depth increases significantly.
-
Implement project-specific adaptation strategies based on project size, language, and maturity; for instance, tailor newcomer retention efforts more intensely for low-size projects and younger ones (low maturity).
-
Develop specialized training or mentorship programs focused on complex tasks within the Python and TypeScript ecosystems to counteract the observed concentration of newcomer dilution effects in these languages.
Abstract
AI coding agents are rapidly proliferating in open-source software development, and the long-term sustainability of open-source projects is closely tied to the health of their human contributor ecosystems. Yet most existing studies treat human contributors as a static backdrop, and the available evidence on human participation after agent adoption mainly concerns newcomer outcomes. This paper takes the human contributor ecosystem as its dependent variable. To the best of our knowledge, it is the first study to estimate how agent adoption changes the number of distinct humans behind each pull request, and to relate this change to agent pull requests recorded under developer accounts. Using the AIDev dataset, GitHub API scanning, agent-trace screening, and two-stage propensity score matching, we construct a balanced panel of 2,246 treated and 3,798 control repositories from January 2023 to May 2026, and estimate effects with the Sun-Abraham estimator for staggered difference-in-differences, event studies, and HonestDiD sensitivity analysis. AI agent adoption is associated with a decline in human contributor density of 0.019 to 0.048, about 4-10% of its pre-adoption level, depending on the reference month (both p < 0.01); because this estimate tolerates only small violations of parallel trends, we regard it as suggestive evidence. The share of newcomers shows no detectable change (ATT = -0.009, p = 0.28), with the 95% confidence interval ruling out a decrease of more than 2.4 percentage points. Whether the decline in density is associated with agent pull requests recorded under developer accounts remains inconclusive (high minus low share: -0.013, 95% CI [-0.083, 0.057]). These findings are consistent with a pattern of augmentation with dilution: we detect no decrease in the share of newcomers, but each human contributor appears to correspond to more pull requests.
Sources
- Agentic Much? Adoption of Coding Agents on GitHub
- How AI Coding Agents Modify Code: A Large-Scale Study of GitHub Pull Requests
- Will It Survive? Deciphering the Fate of AI Agent-Authored Code in Open Source
- Debt Behind the AI Boom: A Large-Scale Empirical Study of AI-Generated Code in the Wild
- On Autopilot? An Empirical Study of Human-AI Teaming and Review Practices in Open Source
- Where Do AI Coding Agents Fail? An Empirical Study of Failed Agentic Pull Requests in GitHub
- The End of Code Review: Coding Agents Supersede Human Inspection
- The Rise of AI Teammates in Software Engineering (SE) 3.0: How Autonomous Coding Agents Are Reshaping Software Engineering
- Early-Stage Prediction of Review Effort in AI-Generated Pull Requests
- From Industry Claims to Empirical Reality: An Empirical Study of Code Review Agents in Pull Requests
Related papers
- ArkEval: Benchmarking and Evaluating Automated CodeRepair for ArkTS
- TestPrism: Rethinking Test Evaluation Beyond a Single Reference
- Mining Workflow Graphs for Black-Box Boundary Testing of Conversational LLM Agents
- Falsification-Based Verification of LLM-Generated Optimization Models: Sound Test Batteries and Their Detection Limits
- GitSkills: A Dataset of Agent Skills on GitHub
- For Coding Agents, the Repository Matters