⚛️ high-energy experiments

AI Agents Can Already Autonomously Perform Experimental High Energy Physics

This paper demonstrates that AI agents, specifically Claude Code within a "Just Furnish Context" framework, can autonomously execute the entire high energy physics analysis pipeline—from data processing and statistical inference to paper drafting—to produce both established and novel results, suggesting the community must adapt its training and workflows to leverage these capabilities for enhanced scientific discovery.

Original authors: Eric A. Moreno, Samuel Bright-Thonney, Andrzej Novak, Dolores Garcia, Philip Harris

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Eric A. Moreno, Samuel Bright-Thonney, Andrzej Novak, Dolores Garcia, Philip Harris

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a high-energy physics experiment as a massive, years-long construction project. Usually, a team of expert architects (physicists) designs the building, but they spend 90% of their time doing the grunt work: mixing the concrete, laying every single brick, and wiring every outlet by hand. They have to write thousands of lines of code just to process data, often copying and pasting the same instructions they used for other projects. It's tedious, prone to human error, and leaves them little time to actually think about why the building should look the way it does.

This paper argues that we can now hire a team of AI agents to do the heavy lifting of the construction, freeing the human architects to focus on the design and the big picture.

Here is a breakdown of what the researchers did, using simple analogies:

The "Just Furnish Context" (JFC) Framework

The researchers built a system called JFC. Think of it as a highly organized construction site manager who doesn't do the physical work but directs a team of specialized robots.

The Boss (The Orchestrator): You give the system a simple goal, like "Measure the weight of this specific particle using old data from the 1990s." The Boss doesn't know the details; it just breaks the job down into steps.
The Workers (Executor Agents): These are the robots that actually write the code, run the simulations, and crunch the numbers. They don't just follow a pre-written script; they have to figure out how to build the solution from scratch, just like a human would.
The Librarian (Knowledge Retrieval): Before building, the agents go to a digital library. They read thousands of old physics papers to see how similar jobs were done in the past. This ensures they don't reinvent the wheel or make silly mistakes.
The Inspectors (Multi-Agent Review): This is the most unique part. Before the humans ever see the result, a team of AI "inspectors" checks the work.
- One inspector checks the physics logic.
- Another checks if the charts look right.
- Another checks if the math adds up.
- If they find a mistake, they send the work back to the "Worker" to fix it. This cycle repeats until the work is perfect.

The "Blind" Test

In physics, there's a rule called "blinding." You aren't allowed to look at the final answer until you are sure your method is correct, so you don't accidentally tweak your math to get the answer you want.

The AI system followed this rule strictly:

Phase 1: It planned the experiment using only theory and old data.
Phase 2: It ran a "partial unblinding" on just 10% of the data to make sure nothing was broken.
The Human Gate: The system stopped and asked a human: "Is everything ready? Can we look at the final 100% of the data?"
Phase 3: Only after the human said "Yes" did the AI look at the rest of the data and produce the final result.

What Did They Build?

The team tested this system on real, open data from famous particle colliders (LEP and CMS). They asked the AI to perform six different complex physics measurements.

The "Reproduction" Test: They asked the AI to recreate a known measurement of the Higgs boson (a famous particle). The AI did it successfully, producing a result that matched the original human-made paper almost perfectly.
The "Discovery" Test: They asked the AI to measure something that had never been measured before in that specific type of experiment (the "Lund jet plane" density). The AI successfully designed the experiment, ran the analysis, and produced a new scientific result entirely on its own.

The Results and the Catch

The paper claims that these AI agents can now do the "boring" part of physics: writing the code, checking the math, and drawing the graphs.

The Good News: The AI produced reports and graphs that looked just like they were written by a junior graduate student. It saved a massive amount of time.
The Reality Check: The authors are very careful to say AI is not replacing physicists.
- The AI is like a very fast, very diligent intern. It can do the work, but it can also make subtle mistakes (like using the wrong formula or missing a tiny detail).
- The human physicist must still act as the "Senior Architect." They have to read the AI's work, understand it, and sign off on it. If the AI makes a mistake, the human is still responsible for it.
- The AI is great at following established rules, but it's not yet good at inventing brand-new, creative ways to solve problems that have never been seen before.

The Bottom Line

This paper shows that we have reached a point where AI can autonomously handle the technical, repetitive side of high-energy physics. Instead of spending years learning how to lay every brick, future physicists can spend their time deciding what to build and why it matters. The AI does the heavy lifting, but the human keeps the keys to the building.

Technical Summary: AI Agents Can Already Autonomously Perform Experimental High Energy Physics

Problem Statement
Experimental High Energy Physics (HEP) analyses are traditionally labor-intensive, algorithmic processes that consume the majority of a graduate student's or postdoc's time. The workflow involves identifying measurement channels, studying literature, designing event selections, quantifying systematic uncertainties, performing statistical inference, and drafting reports. A significant portion of this effort is dedicated to writing thousands of lines of structurally similar code to process data and apply standard techniques, often without deep physics reasoning. This process is frequently slowed by a lack of centralized, up-to-date documentation for core software infrastructure, leading to error-prone and tedious workflows that offer diminishing educational returns after the initial learning phase. While Large Language Models (LLMs) have shown promise in code generation, existing agentic systems in HEP are often limited to rigid, pre-defined analysis structures or require constant human-in-the-loop supervision, preventing true autonomy.

Methodology: The Just Furnish Context (JFC) Framework
The authors introduce "Just Furnish Context" (JFC), a proof-of-concept framework designed to enable fully autonomous execution of HEP analysis pipelines. The framework operates on the principle that an AI agent should receive only a high-level physics objective (e.g., "Measure the Z lineshape") and autonomously determine the full strategy, including event selection, background estimation, and statistical methods, without being provided a template or skeleton code.

Key components of the JFC architecture include:

Orchestrator and Subagents: A central orchestrator delegates tasks to specialized subagents (Executor, Reviewers, Note-writer, Typesetter, Arbiter) using the Claude Code environment (Opus 4.8 model). The orchestrator never writes code or reads full data files; it manages the workflow through written artifacts.
Sequential Phases with Gates: The analysis is decomposed into sequential phases (Strategy, Exploration, Processing, Inference, Documentation). Each phase is gated by the production of a written artifact (e.g., STRATEGY.md, SELECTION.md) and must pass an independent review before proceeding.
Multi-Agent Review System: To replace the distributed human review process of traditional collaborations, JFC employs a panel of specialized AI reviewers:
- Physics Reviewer: Evaluates physics soundness (signal modeling, background estimation) without access to the framework's internal methodology, acting as an external referee.
- Critical Reviewer: Checks compliance with methodology specifications and conventions.
- Constructive Reviewer: Suggests improvements and alternatives.
- Plot Validator: Performs programmatic checks on figures (yields, units, consistency) and detects red flags (e.g., negative yields, non-converged fits).
- Arbiter: Synthesizes reviewer findings into a PASS, ITERATE, or ESCALATE verdict.
Knowledge Retrieval: The framework integrates SciTreeRAG to retrieve domain-specific knowledge from a corpus of published papers (e.g., ALEPH, DELPHI, CMS). This allows agents to adapt selection criteria and systematic uncertainty evaluations based on established experimental practices rather than relying solely on pre-training data.
Blinding Protocol: The framework enforces a strict blinding protocol. Analysis proceeds using Asimov datasets (expected results) until a "Partial Unblinding" phase (10% of signal data) is reviewed and explicitly approved by a human gate. Only after human approval does the agent proceed to "Full Unblinding" on the complete dataset.
Software Stack: The system enforces a pure-Python stack (uproot, awkward-array, hist, pyhf, mplhep) to ensure reproducibility and avoid legacy C++ macro dependencies.

Key Contributions

Autonomous End-to-End Execution: The paper demonstrates that an AI agent can autonomously plan, execute, and document a complete HEP analysis—from event selection and background estimation to statistical inference and paper drafting—based solely on a short natural-language prompt and access to open data.
Novel Measurement: The framework produced the first measurement of the primary Lund jet plane density in hadronic Z decays using ALEPH data. To the authors' knowledge, this is the first time this observable has been measured in $e^+e^-$ collisions and the first new HEP measurement produced entirely autonomously by an AI agent.
Reproduction of Established Results: The system successfully reproduced a CMS Higgs boson signal strength measurement ( $H \to \tau^+\tau^-$ ) in the $\mu\tau_h$ channel using CMS Open Data. The result ( $\hat{\mu} = 1.20 \pm 1.13$ ) is consistent with the published CMS value ( $\mu = 1.01 \pm 0.41$ ) within 0.16 $\sigma$ , validating the framework's ability to replicate complex statistical models.
Multi-Agent Review Validation: The paper establishes that a multi-agent review system can effectively catch errors (e.g., incorrect background priors, over-stated significance) and enforce rigorous standards comparable to human collaboration reviews, reducing the need for continuous human intervention during the coding and debugging phases.

Results
The authors applied JFC to six distinct analyses across ALEPH, DELPHI, and CMS open data:

Scalar Observables: The agent produced measurements for Z lineshape parameters ( $M_Z, \Gamma_Z, \sigma_{had}^0$ ), heavy-flavor observables ( $R_b, A_{FB}^b$ ), and event shapes ( $\alpha_s$ ). Eight of nine scalar comparisons fell within $|pull| < 2$ of published reference values. One outlier ( $\Gamma_Z$ ) showed a $-3.4\sigma$ deviation, traced to a residual efficiency issue at off-peak scan points.
Uncertainty Quantification: While the central values were generally consistent, the agent's uncertainties were often larger than published results (e.g., the CMS $\mu\tau_h$ uncertainty was nearly three times the published value), reflecting the limitations of using partial open datasets and single final states rather than full collaboration datasets.
Qualitative Assessment: The agent consistently selected standard analysis strategies (e.g., iterative Bayesian unfolding, HistFactory likelihoods) and identified correct systematic uncertainty sources. The generated analysis notes and figures were of publication quality, though some required iterative refinement by the review agents to correct formatting or logical errors.
Efficiency: The full pipeline for a single analysis took approximately 10 hours of wall-clock time (including idle time for rate limits), with the active compute time being significantly shorter.

Significance and Claims
The paper argues that the experimental HEP community is underestimating the current capabilities of agentic AI systems. The authors claim that:

Current Capability: AI agents can already handle the "repetitive technical burden" of analysis code development, freeing researchers to focus on physics insight, novel method development, and rigorous validation.
Not Replacement, but Offloading: The authors explicitly state they are not calling for AI to replace physicists. The final output of any AI-assisted analysis must be thoroughly checked and validated by domain experts. The role of the physicist shifts from implementer to architect and critic.
Paradigm Shift: The framework suggests a need to rethink how students are trained, how analysis efforts are organized, and how human expertise is allocated. If implementation can be compressed from years to months (or days), the bottleneck in HEP shifts from coding capacity to physics ideas and human review bandwidth.
Legacy Data Reanalysis: The framework offers a pathway to systematically reanalyze vast archives of legacy data (e.g., LEP, Tevatron) using modern software and techniques, a task previously impractical due to the loss of expertise in legacy codebases (e.g., FORTRAN, PAW).

The paper concludes that while current agents are not perfect and require human oversight for validation and unblinding, the technology to perform graduate-student-level analysis autonomously is already here, warranting immediate experimentation and the development of responsible workflows and benchmarks.