Skip to content

Agent Observatory

Agent Observatory is the provider-free policy workbench included with chemworld lab. It is for understanding how an Agent acts through the public ChemWorld contract—not for displaying private reasoning or claiming leaderboard performance.

Open the public Agent Observatory Open the Student Lab

The free preview may take a short time to wake after an idle period. For a private local instance:

chemworld lab

Open http://127.0.0.1:8876/agent/. The adjacent /student/ route remains the manual workbench.

What is connected

The browser starts the official run_agent() execution path. Every policy therefore receives the same public decision context as a normal CLI run, and every request goes through the normal action validator and transactional environment. The initial catalog includes deterministic reference and replay policies, seeded design and search baselines, and safety-constrained Bayesian optimization. No entry calls an online model.

The Observatory does not fabricate an Agent rationale. It displays only structured trace fields actually emitted by that Agent, alongside the runner's independent decision audit.

Run controls and replay

  • Single step grants exactly one action, then pauses at the next public step boundary.
  • Run continuously continues to task termination; pause takes effect at a step boundary.
  • The timeline can revisit any committed or rejected action without re-running the environment.
  • Download JSON exports the public Observatory projection. Use the CLI JSONL artifact for formal replay verification and evaluation.

Spectral plots are downsampled for responsive rendering. Their source is the public instrument signal returned by the environment; they do not reconstruct hidden chemical state.

Exploratory comparison

The comparison panel runs two to four provider-free policies on exactly the same task and seed and shows endpoint score, cost, safety risk and step count. A single-seed comparison is useful for debugging behavior but is not evidence that one policy is generally superior. Formal comparisons must fix task versions, policy versions, seeds, method-resource limits and evaluation protocol.

Why online providers are not one click

ChemWorld supports custom Python Agents and optional DeepSeek and Codex adapters. They stay in the explicit Python API because live runs require credential handling, model identity, prompt-contract provenance, retry policy and resource limits. See Connect an agent for those paths.

The local web server binds only to loopback and sends no data to an external service. The hosted preview runs the same provider-free catalog behind explicit capacity, expiry, concurrency and rate limits. Both modes store runs only in memory.