thehutch

Observability, steering, and provenance for autonomous-research agents.

thehutch is the PyPI distribution for Hutch — an observability, steering, and provenance dashboard for autonomous-research agents. It covers linear "hypothesis → experiment → claim" pipelines as well as evolutionary, population-based, and self-improving systems (AlphaEvolve / OpenEvolve / ShinkaEvolve / CVEvolve / DGM / SICA / AIDE / ASI-ARCH / FunSearch / POET / MAP-Elites).

Install

pip install thehutch     # PyPI distribution name; imports as `hutch`
hutch serve               # → http://localhost:7777

Three ways to populate the dashboard

(a) Import an existing run. Eleven hand-tuned adapters ship in this release: OpenEvolve, AIDE, DGM, QDax, ASI-ARCH, FunSearch, CORAL, POET, CVEvolve, ptychi-evolve, ShinkaEvolve. For anything else, the LLM-assisted importer asks an LLM to write an adapter on the fly:

hutch import ./checkpoints/circle_packing       # autodetect (11 adapters)
hutch watch ./checkpoints/live_run              # poll and update live
hutch watch ./cvevolve-run --include-audit      # opt into raw CVEvolve audit logs
hutch import ./novel-format --llm                # LLM-assisted fallback

(b) Instrument a Python loop.

import hutch as h
h.start_run(name="my-search")
seed = h.log_individual(kind="hypothesis")
h.log_fitness(individual=seed, scores={"plausibility": 0.7})
h.end_run()

(c) Drop the skill into an LLM-driven agent. See the skill — it makes structured tool calls so any Claude / GPT-4 agent emits canonical events as it works.

Steering

The dashboard can be a control surface for runs that declare capabilities={"steering": True}. Those agents poll hutch.steering.poll() between iterations; the UI's Steering tab issues commands (pause_run, cancel_individual, fork_from, inject_hint, approve_hitl, …). Imported/offline runs stay read-only.

Documentation

• Concepts
• Schema (auto-generated)
• Distribution
• Adapters
• Steering
• Security
• Publication exports

The schema is additive-only between minor releases: new optional fields and new kind enum values are fine; renaming or removing existing fields is a breaking change.

License

Apache 2.0.