Research agents

Fast triage runner — clones, installs (60s cap), runs (90s cap), produces a verdict in under 3 minutes.

A specialist research assistant pre-loaded with a frozen snapshot of all papers in a semantic cluster. Ask for literature reviews, open-problem mapping, cross-paper comparisons, or "what is the state of the art on X in this cluster?" Prompt-cached for fast multi-turn conversations. One agent per cluster, refreshed when the cluster changes.

Clones a paper's GitHub repo into an isolated sandbox, installs dependencies, runs the experiment with a smoke-test budget, compares extracted metrics against the paper's claimed results, and returns a structured verdict. Supports one recovery attempt on install failure. Handles 7 verdict states: success, partial, fails_install, fails_run, timed_out, no_quickstart, unverifiable.

Audits repo reproducibility readiness — license, README, dep pinning, test coverage, missing artifacts.

Lint + type + dead-code analysis on a paper's repo. Surfaces concerns runtime testing won't catch.

Runs full headline benchmarks at the paper's reported config. Outputs a reproduced-vs-claimed table.

Walks 1-hop and 2-hop citations around a paper. Surfaces influential descendants and idea migrations across fields.

Reads across a cluster and ranks the open questions worth attacking next.

Reads a paper's GitHub repository in a sandboxed environment. Checks static indicators (license, README, CI config, dependency pinning, test coverage), optionally runs a quickstart probe, and produces a structured reproducibility verdict with per-claim evidence and red flags. Powered by Claude Sonnet for code-level reasoning.

Cross-references methodology claims against the paper's GitHub repository. Flags missing artifacts, mismatched configs, and runtime claims that cannot be verified from the released code.

Reads a paper's methodology section and identifies design choices, threats to validity, and unstated assumptions. Hands findings to the Reproducibility Scout for runtime verification.

Surveys a research area: identifies key papers, traces methodological lineages, and surfaces open problems. Outputs a structured map of the literature.