Stage 01 / GOAL
Goal before motion.
Harness starts by defining what the agent is allowed to solve, what it must preserve, and how completion will be verified.
AI agent operating contract
Reliability is a system.
Harness gives Claude Code, Codex, and OpenCode a repo-local workflow for goals, context, task boundaries, verification, recovery, memory, and cache-aware routing.
Thin startup cache read98.7%
Warm /wf cache read99.1%
/wf cache-read uplift+5.8 pts
System choreography
One operating loop, staged in motion.
The workflow scales from outline to evidence as the camera moves through the system.
From system to evidence
Evidence is the interface.
Harness makes verification part of the operating contract, not an afterthought.
Why teams use it.
Harness is infrastructure for agent reliability. It does not claim magic productivity numbers; it creates the evidence needed to measure real progress.
Safer changes
Conflict handling classifies create, skip, backup, overwrite, and merge decisions before writes happen.
Recoverable work
Progress files, task capsules, and durable memory make interrupted sessions resumable without re-reading the whole repository.
Verified completion
Tests, validators, review, and browser evidence become part of the definition of done instead of a hopeful final message.
Agent-first install
One prompt. The agent handles the merge.
Give this to the coding agent inside an existing repository. It will choose the safe path, preserve project-owned files, run validation, and report exactly what changed.