MadAgents

arXiv:2601.21015 · hep-ph · Submitted 2026-01-28 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: High Energy Physics: Phenomenology. The day in particle phenomenology: models, signatures and constraints.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "MadAgents".

Mira: Here's the gist of the paper. We present MADAGENTS, a physics-specific agentic tool designed to significantly accelerate LHC research by providing comprehensive support for state-of-the-art simulation tools like MADGRAPH.

Kai: The system encompasses agentic installation, learning-by-doing training for inexperienced users, support for advanced users tackling complex theoretical tasks, and the automation of entire simulation campaigns.

Mira: The core challenge in the precisionLHC program is analyzing vast amounts of data in quantum field theory.

Kai: This relies on comparing measured scattering events with first-principle predictions from event generators (e.g., PYTHIA, MADGRAPH, SHERPA).

Title and authors: Kai: So we're looking at this paper called MadAgents, and it’s about this agentic tool designed to help with simulation work in particle physics. It sounds like they're trying to make the whole process of using tools like MADGRAPH much faster for LHC research.

Mira: Right, it seems the main idea here is building an AI system that can handle installing software, training people on how to use it, and even running whole simulation campaigns automatically from a document. It’s about bridging the gap between complex quantum field theory predictions and the massive amounts of data we collect at the LHC.

Lev: From my side, I'm interested in how this system actually handles the heavy lifting of simulation. If it can automate event generation and run campaigns autonomously, that means we could spend less time manually setting up things like those complex NLO calculations and more time actually looking for new physics signals on the data.

Kai: Exactly. The paper shows they built a system with these specialized agents, like an orchestrator and different worker agents for specific tasks—things like setting up MADGRAPH or writing bash scripts. It’s not just one big program; it’s a whole team of AI components working together.

Mira: And what I find interesting is their focus on how this acceleration applies to the actual users, whether they are beginners who need tutorials, or advanced users tackling really hard theoretical problems involving things like off-shell effects. It’s tailored support for different skill levels.

Lev: That tailoring is key because running a full simulation campaign autonomously requires a very precise workflow that an experienced user might know by heart but a novice wouldn't grasp the setup for. The paper suggests this agentic approach helps manage that complexity.

Kai: And they also talk about this self-improvement loop they built using Claude Code, where the AI tests its own knowledge against specific MADGRAPH topics and corrects documentation errors itself. That’s a neat way to make the tool more robust over time.

Mira: That sounds like a self-correcting mechanism, which is important when you're dealing with so much complex theory and potential for human error in setup scripts or formulas. It moves the system beyond just having static knowledge.

Lev: If it’s correcting documentation about things like NLO compatibility warnings or reweighting formulas for MadSpin modes, that shows they are actually digging into the technical details of how these simulation tools behave under pressure. That kind of correction is valuable for real-world hardware runs too.

Title and authors: Kai: So, to wrap up on what they did: this paper introduces MadAgents as a system that offers agentic installation, learning-by-doing training, and autonomous simulation campaigns. It’s designed to make state of the art simulations like MADGRAPH easier and faster for researchers at the LHC.

Mira: And their second part is showing how this system can take a PDF of a paper and run an entire simulation campaign on its own, which is pretty ambitious when you think about the complexity involved in setting up those physics models. It’s trying to automate the research workflow itself.

Lev: The improvements they suggest are also interesting because they focus on making the agents more reliable by adding strict evidence requirements. For example, forcing worker agents to cite official documentation or local software outputs instead of just guessing is a big step toward trustworthiness in an AI assistant.

Kai: That evidence requirement is pretty important because when you’re dealing with precision physics, you can’t just trust an answer; you need to know exactly where it came from. It forces the AI to ground its claims in verifiable facts about the software itself.

Mira: I also like how they refined the agent structure by splitting roles—making a plan planner and a verification reviewer—to make sure there’s a better check on correctness throughout the process, not just at the end. That separation of concerns sounds like it helps manage complexity better.

Lev: And I have to mention that they replaced the plan updating tool with something that avoids an extra large language model call, which makes the entire system run much faster and cheaper for iterative work. Latency matters when you're running hundreds of these kinds of experiments.

Kai: So, it seems like they’re focusing on making the workflow efficient not just for the user learning, but for the machine doing the simulation itself. It’s about optimizing the computational structure as much as we optimize the physics calculations.

Mira: Overall, this paper points toward a future where research workflows are guided by these specialized AI agents that can handle everything from initial setup to post-analysis automation. It shifts the focus from just calculating one thing to managing the entire data analysis pipeline efficiently.

Lev: If you take all of this together, MadAgents aims to significantly accelerate LHC research by allowing autonomous workflows and training physicists at the same time. It’s a tool that tries to handle the entire lifecycle of a simulation task with minimal human intervention.

Title and authors: Kai: So, we’re seeing an effort here to make the complex world of quantum field theory simulations more accessible through these structured agentic tools. It's about providing a comprehensive support layer for cutting edge simulation tools like MADGRAPH seven <ref:2601.21015#pg1>.

Mira: And it’s not just about making it easier for the user; it’s about enabling entirely new ways to explore the data, like automating campaigns based on existing scientific literature. That level of automation is pretty significant for discovery science.

Lev: It really highlights that the computational efficiency of event generators has become a research bottleneck, and tools like this try to address that by being smarter about their own usage and the surrounding analysis steps.

Kai: So, looking ahead, it seems like the next step for these kinds of systems is to see how far they can push this automation into fully independent hypothesis generation or complex multi-step experimental design. It’s a big leap from just running simulations on command to being truly autonomous research partners.

Mira: That's where the real implications lie, I think. If we can automate the process of taking a paper and setting up a simulation campaign, it means researchers could test more hypotheses in less time than they currently spend on setup and initial runs. It’s about scaling the exploration of new physics models.

Lev: From an engineering standpoint, the challenge will be ensuring that this autonomous workflow is robust enough to handle the inevitable unexpected errors that don't show up in standard datasets. The self-improvement loop is a good start, but real hardware demands extreme reliability from the underlying components.

Kai: So, to wrap things up on MadAgents: it’s a comprehensive agentic system that brings installation and training together with autonomous campaign running, all while forcing the agents to use verifiable evidence for their claims. It’s about making complex simulation workflows faster and more trustworthy for LHC research.

Mira: And I think the real shift here is moving toward AI as an integrated part of the research pipeline itself, not just a helpful assistant you ask questions to. It’s about enabling researchers to treat the entire process from literature review to simulation execution as a single, agentic task.

Lev: To sum up: this paper lays out how agentic support for tools like MADGRAPH can drastically speed up LHC research by enabling autonomous workflows and training physicists simultaneously. It’s a tool that tries to handle the entire lifecycle of a simulation task with minimal human intervention, focusing on making the computational structure faster and more reliable.

The paper's summary: Kai: So, we're looking at this paper called MadAgents, and it’s about building an agentic tool for physics simulations at the LHC.

Mira: Right, it’s basically an AI system designed to automate a huge chunk of the work physicists do when they use tools like MADGRAPH.

Kai: It handles everything from installing software itself to training new people on how to use it, and even running whole simulation campaigns just by reading a research paper.

Mira: The big idea is bridging the gap between writing down those first-principle predictions and actually getting that data analysis done without constant manual intervention.

Kai: They’re using these specialized agents—like an orchestrator and different workers for specific tasks—to make the whole workflow happen automatically.

Mira: And they built a self-improving loop where the AI tests its own knowledge against specific physics topics and corrects its own documentation errors.

Kai: That means the system is constantly refining how it understands things like NLO compatibility or reweighting formulas for MadSpin modes.

Mira: It shows they’re moving toward systems that can handle complex research workflows more independently.

Kai: This paper validated it by running a simulation using a hundred thousand events, looking at how different simulation strategies affect the results for top quark pair production.

Mira: The findings show that just picking one way to simulate isn't enough; you actually need to look at the whole matrix-element information from multiple jet choices.

Kai: They found that using NLO plus parton showers on-shell gives you a good shape at low energies, but you still have uncertainty in the tail of the distribution.

Mira: And if you only use leading order off-shell, it smooths out those line shapes, but then the uncertainty in those high energy tails gets really big because of the radiation that's missing.

Kai: The best approach they tested was using leading order merged with parton showers and then adding parton showers again, which gives a more physical shape even if the normalization stays at leading order.

Mira: So for someone just listening, this means a researcher could theoretically take a new paper, feed it into this system, and get a full simulation campaign running without spending hours setting up all the individual pieces.

Kai: It’s about automating the entire research pipeline from literature review to simulation execution and training at once.

Mira: And they are making sure that this automation is trustworthy by forcing the agents to provide actual evidence for their claims, so you know where their answers are coming from.

Kai: This moves us away from just asking an assistant questions and toward having a partner that manages complex tasks autonomously.

Mira: It really highlights how computational efficiency in these simulation tools has become a major bottleneck in making discoveries at the LHC.

Kai: The next step for this kind of system is seeing how far they can push this automation into generating new hypotheses on its own, not just running simulations based on existing ones.

The paper's improvements: Tom: So, we're looking at how they plan to make MadAgents even better, and they’re focusing on making sure the agents are actually reliable when used in a real research environment.

Kai: The main thing is adding this strict evidence requirement for all claims made by the workers.

Mira: That means before an agent can state something as verified, it has to show you proof from official documentation or by running a local software command and showing the exact output.

Kai: It’s forcing them to ground their answers in actual facts about MADGRAPH or other tools instead of just relying on their internal knowledge.

Mira: That also means they're refining the structure of the agents themselves, splitting up roles like planning, verifying correctness, and presenting results to catch errors earlier.

Kai: That separation is smart because it lets each agent focus tightly on one job—planning versus checking if the plan is physically sound versus just formatting the final output.

Mira: And they’ve also made a big change by replacing the plan updating tool with something that doesn't need an extra large language model call every time.

Kai: That should really help cut down on latency and costs when you’re running these complex workflows repeatedly.

Mira: It sounds like they are trying to make the whole system faster and more cost-effective for iterative research work, which is something that matters a lot in this field.

Kai: I think it shows they realized that the complexity of managing agents wasn't just about making them smarter, but also making the entire management process itself more efficient.

Mira: If they can handle planning and verification with fewer steps and less computational overhead, then we could see these agentic tools being used more frequently in actual data analysis pipelines.

Kai: That would mean researchers spend less time debugging the setup of their simulations and more time looking for new signals in the data itself.

Mira: It shifts the focus from just making one simulation run work to managing an entire automated research workflow that is both fast and verifiable.

Kai: So, they’re not just building a smarter assistant; they’re building a more robust system designed to operate reliably in the messy reality of particle physics experiments.

Mira: And this reliability is key because when you're dealing with high precision measurements at the LHC, you need confidence in what your simulation tools are actually doing.

Kai: The next thing we should look into is how this whole agentic structure might interface with actual machine learning models for data extraction and analysis.

Conclusion: Tom: So, to wrap up this discussion on MadAgents, we're summarizing how this agentic tool aims to change LHC research workflows.

Kai: Basically, MadAgents is an AI system that automates everything from setting up simulation software like MADGRAPH to running entire complex simulation campaigns based on a paper.

Mira: It’s designed for researchers so they can focus less on the tedious setup and more on finding new physics signals in the data.

Kai: The core innovation is using these specialized agents that can learn by doing, providing support for both beginners and people tackling really deep theoretical problems.

Mira: And they built a self-improving loop where the AI checks its own understanding against specific physics topics and corrects its documentation errors on the fly.

Kai: That means the system is constantly refining how it understands things like NLO calculations or reweighting formulas for those MadSpin modes we talked about earlier.

Mira: It shows they’re moving toward systems that can handle complex research workflows more independently, which is a big step.

Kai: This paper validated it with a simulation of a hundred thousand events, showing how different simulation strategies affect the final result for top quark pair production.

Mira: They found that you need to look at the whole picture—like merged matrix-element information—instead of just choosing one simple kinematic setting.

Kai: It’s about making the computational structure faster and more reliable so researchers can spend less time debugging setups and more time looking for new physics.

Mira: This paper points toward a future where AI is integrated into the research pipeline itself, handling everything from literature review to simulation execution as one task.

Kai: So, this whole MadAgents system aims to significantly accelerate LHC research by enabling autonomous workflows and training physicists simultaneously.

Mira: It’s about scaling the exploration of new physics models by automating the setup phase of experiments.

Lev: From my side, I see a lot of value in that evidence requirement they added for claims; it forces the agents to cite actual software output instead of just making up answers.

Kai: That grounding is important because when you're dealing with precision physics, you can’t just trust an answer without knowing exactly where it came from.

Mira: And I think the next big step for these tools is seeing how far they can push this automation into generating entirely new hypotheses on its own, not just running simulations based on existing ones.

Lev: If we want to run this kind of autonomous workflow on real hardware, the challenge will be ensuring that it stays reliable when it hits those unexpected errors that don't show up in standard datasets.

Kai: So, to wrap things up on MadAgents, it’s a comprehensive system for installation and training that brings autonomous campaign running together.

Mira: It really shifts the focus from just calculating one thing to managing the entire data analysis pipeline efficiently.

Lev: We’re looking at how much overhead this kind of agentic management would take when we try to run it on actual computing resources, and that's where the engineering questions come in.

Institut für Theoretische Physik, University of Heidelberg · Interdisciplinary Center for Scientific Computing, University of Heidelberg

hep-ph

Submitted: 2026-01-28

Updated: 2026-09-11

Comments: 61 pages, 3 figures, 1 table. v3: includes a Claude Code implementation with a self-improvement loop

Journal ref: SciPost Phys. Codebases 86 (2026)

DOI: 10.21468/SciPostPhysCodeb.86

Code: https://github.com/MadGraphTeam/MadAgents

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 92/100

The gist: We present MADAGENTS, a physics-specific agentic tool designed to significantly accelerate LHC research by providing comprehensive support for state-of-the-art simulation tools like MADGRAPH.

Key concepts

Agentic System
A sophisticated AI structure using multiple specialized agents (like an Orchestrator and a MG-Operator) that work together to complete complex tasks. These agents can plan, execute specific functions, and review each other's work to achieve goals autonomously.
Evidence Requirement
A strict rule forcing worker agents to cite verifiable proof—such as official documentation or exact software output—before making a claim. This mechanism prevents incorrect statements by grounding the agent's responses in concrete, traceable data rather than general knowledge.
Self-Improvement Loop
A continuous process where the system tests its own understanding of MADGRAPH topics by generating questions and evaluating its answers against curated documentation. This loop identifies and corrects errors in internal knowledge, making the agent more robust over time.
Simulation Strategy
The specific methods used to generate physical simulation data, such as choosing between NLO+PS on-shell versus LO merged off-shell calculations. The paper shows that different choices affect the resulting data shape and uncertainty in high-energy physics predictions.

Terminology

Summary

We present MADAGENTS, a physics-specific agentic tool designed to significantly accelerate LHC research by providing comprehensive support for state-of-the-art simulation tools like MADGRAPH. p. 0 The system encompasses agentic installation, learning-by-doing training for inexperienced users, support for advanced users tackling complex theoretical tasks, and the automation of entire simulation campaigns. p. 0

The core challenge in the precisionLHC program is analyzing vast amounts of data in quantum field theory. p. 2 This relies on comparing measured scattering events with first-principle predictions from event generators (e. p. 2g., PYTHIA, MADGRAPH, SHERPA). p. 2 Modern machine learning (ML) allows for extracting information from these large datasets, but this acceleration must also cover the efficiency of simulation tools and analysis workflows. p. 2 MADAGENTS addresses this by providing ML-accelerated user training and support tailored to the next-generation computational structure of MADGRAPH 7. p. 2

MADAGENTS functions as a sophisticated agentic system, utilizing LangGraph with specialized controlling agents (Orchestrator, Planner, Reviewer) and worker agents (MG-Operator for MADGRAPH specialization, Script-Operator for bash/PYTHON specialization, CLI-Operator for user access, Plotter for figure creation, PDF-Reader for document extraction). p. 4 A key feature is the addition of evidence requirements to prevent incorrect statements by demanding supporting evidence from official sources or local software invocations.

The capabilities demonstrated include:

  1. Installation and Setup: MADAGENTS can autonomously install MADGRAPH along with related codes like DELPHES and ROOT, including building ROOT from source.

  2. Inexperienced User Support: It designs dedicated MADGRAPH tutorials, runs simulations, and performs standard tasks such as event reweighting for newcomers.

  3. Advanced User Support: It assists experienced users in complex tasks, such as simulating top pair production beyond leading-order on-shell precision and understanding underlying theory assumptions (e.g., NLO+PS vs off-shell effects). p. 2

  4. Autonomous Simulation Campaigns: MADAGENTS can reproduce entire simulation campaigns from publications without user intervention, extracting relevant information autonomously from articles and web content.

The system's self-improvement mechanism is implemented via a Claude Code implementation featuring a self-improvement loop across six phases (Question generation, Evaluating MADAGENTS, Extracting and validating claims, Grading and diagnosing, Documentation improvement, and Re-evaluation). This loop tests the agent's understanding against specific MADGRAPH topics by curating internal documentation through repeated questioning with Opus 4. This process has successfully identified and corrected errors in the documentation regarding NLO compatibility warnings or incorrect reweighting formulas for MadSpin modes, shifting the focus from built-in knowledge to robust reasoning capabilities.

The system is validated using a controlled dataset simulating effective-field-theory (EFT) top–antitop production at 13 TeV, including detector effects at the Delphes level. p. 2 This dataset contains N=100,000 events generated with MadGraph5_aMC@NLO interfaced with Pythia8 and Delphes 3 for fast detector simulation. The analysis focuses on studying EFT effects in pp to t t bar production with a non-zero chromomagnetic operator (ct G). p. 4

Key findings regarding simulation strategies for top pair production highlight the importance of matrix-element multi-jet information (merged sample) over just kinematic choices:

  • Using NLO+PS on-shell yields correct inclusive normalization and decent shape at low–moderate p T, but residual modeling uncertainty in the tail.

  • Using only LO+PS off-shell without merging improves off-shell line shapes but suffers from large uncertainties in the p T tail due to LO radiation.

  • Using LO merged t t bar (+0, 1j) + PS provides a more physical shape where a single hard extra jet dominates the recoil, although normalization remains LO.

The study concludes that agentic support for tools like MADGRAPH significantly accelerates LHC research by enabling autonomous workflows and training physicists simultaneously. p. 0

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

  1. Evidence requirement for claims is strictly enforced by adding explicit instructions to worker agents: Never base your answers on your general knowledge of MadGraph or related tools; worker agents must present supporting evidence. This forces agents to cite official/trustworthy documentation, a local software invocation (exact command + exact output), or a relevant excerpt from the software source code to label claims as VERIFIED.

  2. Improved agent structure separates concerns by refining the Reviewer role into Plan-Reviewer, Verification-Reviewer, and PresentationReviewer for better error detection. This allows agents to focus on planning, correctness, and presentation, leading to more robust workflow management.

  3. The Plan-Updater agent is replaced by a plan updating tool that avoids an unnecessary LLM invocation. This reduces latency and cost associated with the Orchestrator's frequent requests for structured updates, making the overall system faster and more efficient.

  4. Worker agents are given their own dedicated conversations, allowing them to recall previous chain-of-thought and tool interactions without confusion from the Orchestrator's messages disguised as user communication. This ensures workers maintain context relevant to their specific task execution.

  5. The Orchestrator is equipped with a scratchpad for short-term notes and can now issue tool calls while preserving the chain of thought, enabling it to dispatch multiple agents in parallel for complex workflows. This allows for more sophisticated, multi-agent orchestration than previously possible.

Sources

Related papers