Rollout, Replay & Forensics
Every agent run that has rollout recording enabled writes a JSONL file — called the rollout — that captures every event emitted during that run. That file becomes your entry point for three progressively richer forensics capabilities:
Training & Benchmarks
The training and benchmark system gives you two ways to measure and improve agent quality. trainerd is a Go CLI that scores and analyzes run traces produced by the harness — without ever re-running the agent. The benchmark harnesses (a key-free Go smoke, a Terminal-Bench integration, an overnight loop, and a comparison harness) let you measure task-completion accuracy across real workloads. Together they form a feedback loop: run tasks, capture rollout files, score them, find anti-patterns, apply fixes to prompts or tool descriptions, and guard against regressions.
Development & Testing
go-code is developed and tested without any LLM API key for the vast majority of the test surface. This page explains how to bootstrap a contributor development environment, write and run deterministic tests using the fake provider, exercise the full regression suite, and understand how CI gates work.