AI Agents Can Already Autonomously Perform Experimental High Energy Physics
| Authors | Eric A. Moreno et al. |
| Year | 2026 |
| Field | AI / Agents |
| arXiv | 2603.20179 |
| Download | |
| Categories | cs.AI, cs.LG |
Abstract
Large language model-based AI agents are now able to autonomously execute substantial portions of a high energy physics (HEP) analysis pipeline with minimal expert-curated input. Given access to a HEP dataset, an execution framework, and a corpus of prior experimental literature, we find that Claude Code succeeds in automating all stages of a typical analysis: event selection, background estimation, uncertainty quantification, statistical inference, and paper drafting. We argue that the experimental HEP community is underestimating the current capabilities of these systems, and that most proposed agentic workflows are too narrowly scoped or scaffolded to specific analysis structures. We present a proof-of-concept framework, Just Furnish Context (JFC), that integrates autonomous analysis agents with literature-based knowledge retrieval and multi-agent review, and show that this is sufficient to plan, execute, and document a credible high energy physics analysis. We demonstrate this by conducting analyses on open data from ALEPH, DELPHI, and CMS to perform electroweak, QCD, and Higgs boson measurements. Rather than replacing physicists, these tools promise to offload the repetitive technical burden of analysis code development, freeing researchers to focus on physics insight, truly novel method development, and rigorous validation. Given these developments, we advocate for new strategies for how the community trains students, organizes analysis efforts, and allocates human expertise.
Engineering Breakdown
Plain English
This paper demonstrates that Claude Code, a large language model-based AI agent, can autonomously execute the entire high energy physics (HEP) analysis pipeline with minimal human guidance. The researchers gave the agent access to HEP datasets, an execution framework, and prior experimental literature, and found it successfully completed all stages: event selection, background estimation, uncertainty quantification, statistical inference, and even paper drafting. The key finding is that current LLM agents are significantly more capable than the community realizes, and existing agentic workflows are often over-constrained by excessive scaffolding. The authors present Just Furnish Context (JFC), a minimal-scaffolding framework that enables autonomous execution of complex scientific analysis workflows.
Core Technical Contribution
The core novelty is demonstrating that general-purpose LLM agents can autonomously navigate and execute domain-specific scientific analysis pipelines without task-specific fine-tuning or narrowly-designed scaffolding. Rather than building rigid workflow systems that constrain the agent to predefined steps, the authors show that simply providing context (datasets, tools, literature) allows the agent to discover and execute the correct analysis sequence. The JFC framework operationalizes this insight: it minimizes hand-crafted structure while maximizing the agent's autonomy and reasoning capability. This is a significant shift from prior agentic work that relied heavily on prompt engineering, chain-of-thought scaffolding, and explicit state machines to guide agent behavior.
How It Works
The system starts by ingesting a HEP dataset, domain-specific tools (data loaders, statistical libraries), and a corpus of prior published HEP analysis papers as context. Claude Code, an LLM with code execution capabilities, receives this context plus a high-level goal (e.g., 'perform a complete HEP analysis on this dataset'). The agent generates Python code iteratively to explore the data, formulate hypotheses about signal versus background processes, and implement statistical tests. At each step, the agent executes code, observes results, and decides next steps based on the output—this closes the loop between reasoning and execution. The agent handles complex tasks like setting selection cuts to maximize signal significance, estimating backgrounds from data sidebands, propagating systematic uncertainties through a likelihood fit, and computing p-values and confidence intervals. Finally, the agent synthesizes findings into a written analysis summary or draft paper section, demonstrating end-to-end capability across the full scientific workflow.
Production Impact
This work directly impacts how scientific and analytical organizations can accelerate exploratory analysis and reduce the human expertise bottleneck. In practice, domain teams could feed raw datasets and tool libraries to an agent, dramatically shortening the time from data acquisition to publishable results—potentially cutting analysis timelines from weeks to days for standard workflows. However, production deployment requires careful integration: you need robust execution sandboxing (untrusted code generation from LLMs is a security risk), version control for generated analysis code, human review checkpoints for critical decisions (significance thresholds, systematic assignments), and clear audit trails for reproducibility and regulatory compliance. Compute costs are non-trivial—interactive multi-turn reasoning with code execution and large context windows (literature corpus) will require significant API calls or on-premise inference infrastructure. The approach also assumes tool availability: the agent needs well-documented, composable libraries; if your domain lacks this, you'll spend effort building tool abstractions before the agent becomes useful.
Limitations and When Not to Use This
The paper does not address failure modes when the agent encounters novel or out-of-distribution data patterns; it's unclear how gracefully the system falls back to human intervention when uncertainty is high. The approach assumes comprehensive prior literature is available and accurate; in emerging research areas with sparse or contradictory prior work, the agent may propagate biases from the training corpus. Systematic uncertainties and domain assumptions (e.g., which backgrounds to consider, how to model detector effects) are implicit in the provided tools and context—the agent is not creating new physics insights, merely executing known workflows more autonomously. The paper lacks quantitative evaluation metrics; there's no comparison of agent-generated analyses to human expert analyses on identical datasets, so it's unknown whether the agent's independence comes with a quality trade-off. Additionally, legal and ethical liability is unresolved: if an agent-generated analysis leads to incorrect scientific conclusions or business decisions, who bears responsibility?
Research Context
This work builds on the recent surge in LLM-based agents (ReAct, Toolformer, Code Interpreter) that couple language models with tool use and feedback loops, but applies these primitives to real-world scientific workflows at scale. It challenges the prevailing agentic design philosophy, which emphasizes explicit task decomposition and prompt scaffolding (e.g., few-shot examples, chain-of-thought templates); the authors argue this over-engineering limits agent autonomy and generalization. The paper indirectly relates to research in AI-assisted scientific discovery and autonomous systems, though most prior work focuses on hypothesis generation or experiment design rather than full end-to-end analysis pipelines. It opens a new direction: auditing whether domain-specific benchmarks and datasets are sufficient for autonomous execution, and whether the bottleneck to agent capability is technical (model capacity) or organizational (tool/context availability).
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::