Key Points

Generative AI as an agentic research tool

A conversational model returns text; an agentic harness executes tools and conditions on their output.
Level 1, the agent loop, runs in a harness with four parts: the model, the context window, a set of typed tools, and a control loop — extended by MCP tools, subagents, skills, hooks, and monitors.
Level 2 wraps the agent in a verification loop: a grader checks each result against explicit success criteria, because stochastic output must be treated as a hypothesis, not an answer.
Level 3 runs the verified agent on events and schedules; level 4 feeds what actually happened back into the harness, where gains compound.
Building on the open Model Context Protocol keeps tools portable across assistants and supports reproducibility.

The observable is the p π⁻ invariant mass; the Λ⁰ appears as a narrow peak over a combinatorial background.
The peak width is set by detector resolution, not the negligible Λ⁰ natural width.
The data are EDM4eic collections in an events tree; momenta are in GeV.

MCP is a JSON-RPC client–server protocol; a server exposes tools, resources, and prompts to any compliant client.
The uproot server returns compact, JSON-serializable summaries rather than raw arrays, which keeps results inspectable.
execute_kernel runs client-supplied Python in a restricted sandbox: no imports or I/O, only NumPy/awkward, with a timeout.
The servers run inside eic-shell (eic-mcp up) and speak streamable HTTP; opencode and other clients connect to the same 127.0.0.1 URLs (eic-mcp config <client>).
PODIO/uproot is one access path; RDataFrame, TTreeReader, and bare uproot give the same result (see the extras).

AGENTS.md is always-loaded project context; a SKILL.md is a named procedure loaded on demand.
Keep one source of truth (AGENTS.md) and point tool-specific files (copilot-instructions.md, .cursorrules) at it — never maintain duplicates.
A skill’s frontmatter description is what the model matches against to decide when to load it.
Encode inputs, steps, success criteria, and provenance so a result can be reproduced and audited.

The full analysis combines Episodes 1, 3, and 4: the agentic loop, the uproot MCP tools, and the lambda-fit skill.
The same kernel scales from execute_kernel (one root:// file) through execute_kernel_dataset (a small batch) to the async submit_kernel_dataset (the full sample), merging into one histogram.
The yield comes from a Gaussian-plus-polynomial fit; report \(\mu\), \(\sigma\), \(S\), and \(\chi^2/\text{ndf}\), not a bare count.
Pinning inputs and recording tool calls make the measurement reproducible and auditable.

The EIC exposes its infrastructure through MCP: analysis (uproot), data (xrootd, rucio), records (zenodo, indico), code (LXR + GitHub), and production (PanDA, PCS) — three of which you ran yourself.
The DISpatcher bot is the zero-setup path: ~100 MCP tools behind a Mattermost account, no client configuration at all.
corun-ai/codoc-ai is the long-latency complement: high-level models, preserved and expert-curated outputs, grounded in nightly-indexed code knowledge.
Patterns to reuse: tiered tool exposure (context economy at scale) and the secret-token fabrication check (verification over confidence, enforced).
This catalog dates quickly — check the eic GitHub organization and the ePIC dev-cloud for the current list.