harness-engineer

Diagnose the last agent session failure and generate harness patches to prevent recurrence. Reads git log, circuit breaker logs, progress files, and feature state.

Garbage collect the harness. Find and fix stale docs, layer violations, stuck features, dead context, and entropy accumulated from agent-generated code.

Scaffold a complete harness for this project. Generates AGENTS.md, features.json, init.sh, claude-progress.txt, layers.json, docs/ structure, and wires circuit breaker hooks.

Explore harness engineering theory — context engineering, Constitutional AI alignment, CLASS evaluation metrics, and anti-patterns. Optionally filter by topic.

| Agent | File | Purpose |

The Orchestrator is the master routing agent for multi-stage harness engineering workflows. When a user presents a request that spans multiple concerns — project setup, failure recovery, maintenance, theory consultation — the Orchestrator classifies the request, selects the right skills, sequences them, and synthesizes their outputs into a unified report.

Analyze agent failures and automatically generate harness fixes to prevent recurrence. Use whenever an agent session went wrong, produced broken output, got stuck in a loop, failed to complete features, or the codebase is in a bad state after an agent run. Reads git log, progress files, circuit breaker logs, and test output to diagnose the failure mode, then generates a targeted harness patch (AGENTS.md update, features.json fix, new hook, or architectural constraint) that prevents the same failure from happening again. Based on LangChain's Trace Analyzer and OpenAI's "failure = harness signal" principle. Trigger on: "agent failed", "session went wrong", "agent got stuck", "loop broke", "agent made a mess", "why did it fail", "harness doctor", "fix my harness".

Expert guide for designing autonomous AI agent systems using harness engineering principles — the full environment of scaffolding, constraints, alignment, and evaluation that makes AI agents production-reliable. Use this skill whenever the user asks about: building AI agent systems or pipelines, designing context/prompt/scaffolding architecture, setting up evaluation frameworks or benchmarks for AI agents, aligning AI behavior with constitutional principles, measuring AI agent performance (accuracy, latency, cost, safety), debugging or improving agent reliability, or any request involving "agents", "harness", "evals", "scaffolding", "LangSmith", "CI/CD for AI", "agentic workflows", "multi-agent systems", or similar.

Garbage collect the harness: find and fix stale documentation, architectural layer violations, inconsistent feature states, dead context, and entropy accumulated from agent-generated code. Run periodically (weekly or after large sprints) to keep the harness healthy and agents oriented correctly. Produces a prioritized cleanup report and applies fixes — all written by the agent, never manually. Based on OpenAI's "garbage collection agents" pattern and the principle that documentation for agents must be maintained by agents. Trigger on: "gc the harness", "clean up harness", "docs are stale", "harness drift", "garbage collect", "harness maintenance", "agent keeps getting confused", "clean up features.json".

Scaffold a complete harness for any new project or repo so AI coding agents can work effectively across multiple sessions. Use whenever starting a new project with Claude Code, Codex, or any AI agent, or when an existing project has no harness and agents keep losing context between sessions. Creates: AGENTS.md table-of-contents, docs/ knowledge base, features.json with all features pre-marked failing, init.sh dev server startup script, claude-progress.txt handoff file, current_tasks/ multi-agent lock directory, layers.json architectural dependency graph, and wires the circuit breaker hooks. Trigger on: "set up harness", "scaffold this project", "init harness", "new project", "agent keeps forgetting context", "set up for agent-first development".

Session startup ritual for AI coding agents — orients a fresh agent at the start of every Claude Code session in a harness-enabled project. Automatically fires via SessionStart hook. Reads progress files, git history, feature list, runs the dev server smoke test, picks the next feature, and primes the agent with the reasoning sandwich (xhigh planning → high implementation → xhigh verification). Eliminates the "new agent arrives with no context" problem described in Anthropic's long-running agent research. Also implements LangChain's LocalContextMiddleware pattern: maps directory structure, available tools, and injects environment context upfront. Trigger on: "start a new session", "orient the agent", "session startup", "onboard", auto-fires at SessionStart in harness-enabled projects.