The control plane for
auto-research work.
arcp orchestrates a fleet of agents — code improvers, paper-to-code pipelines, thesis explorers — over a single API and a single UI. Submit a use case, watch the events stream, review the result.
What you can do
Each card is a research workflow. Behind each one is a job kind, a spec schema, and one or more sibling agents that arcp orchestrates.
Explore & validate a thesis
Map a research space with a search-driven literature scan, then validate the thesis end-to-end — A/B variants with a judge model or benchmark inference for latency and quality. The validation platform, built on the fleet.
Build a repo wiki
autowiki maps a repository into a living wiki — architecture, modules, and a code graph. Deterministic, no LLMs required.
Run E2E research
autoresearch takes a thesis end-to-end: explore, experiment, synthesize, and optionally implement the result.
Improve a codebase
Run autocode against a repo to refactor, add tests, or apply a research recipe. Local by default; toggle on cloud sandboxing when isolation matters.
Paper → code
Hand paper2code a PDF or arXiv id. It paces through extraction, planning, drafting, and verification — pausing for approval at the gates that matter.
Code → paper
code2paper turns a thesis, experiments, and benchmarks into a publishable output — blog, whitepaper, research paper, or executable notebook.
Post-train a model
Point autotune at a base model and dataset and pick a method — SFT, RL, DPO, or LoRA. Get back a trained adapter plus metrics, GPU work routed through Tensile.
Improve a live agent
autoimprove reads a running agent's logs and synthesizes an optimal training loop — rubrics, verifiers, and a dataset — ready for autotune.
How it's wired
arcp
API + jobs + UI. Holds the queue, the journal, and the orchestration logic. The piece you're looking at.
autocode
The code agent. Runs locally by default for low-friction loops. Sandbox mode (Modal / Docker / E2B) is opt-in.
paper2code
Paper → code pipeline as its own service. arcp talks to it over HTTP, the same way it talks to the other sibling agents.
Tensile — the Research Commons infrastructure layer.
Schedulers, sandboxes, observability, and shared services that every RC product runs on. arcp is one of those products.