agent-harness
Turn any domain folder of skills into a bounded agentic loop: compile a goal into a verifiable task plan, execute tasks with the domain's own tools, verify every task with machine-run checks, retry with caps, escalate to a human when budgets exhaust, and refuse to close until everything is verified or explicitly waived. Use when you want an agent or subagent to pick up a goal and drive it to a verified close across one of this repo's 18 domains ('run this goal through the engineering harness', 'set up an agentic loop for marketing work', 'make the finance domain self-verifying'). NOT for authoring Claude Code Workflow-tool .js scripts (workflow-builder), N-agent tournaments on one task (agenthub), single-file metric optimization (autoresearch-agent), or discovering published loop recipes (loop-library).
alirezarezvani · AI review on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add alirezarezvani/claude-skills --skill agent-harnessAbout this skill
name: agent-harness description: "Turn any domain folder of skills into a bounded agentic loop: compile a goal into a verifiable task plan, execute tasks with the domain's own tools, verify every task with machine-run checks, retry with caps, escalate to a human when budgets exhaust, and refuse to close until everything is verified or explicitly waived. Use when you want an agent or subagent to pick up a goal and drive it to a verified close across one of this repo's 18 domains ('run this goal through the engineering harness', 'set up an agentic loop for marketing work', 'make the finance domain self-verifying'). NOT for authoring Claude Code Workflow-tool .js scripts (workflow-builder), N-agent tournaments on one task (agenthub), single-file metric optimization (autoresearch-agent), or discovering published loop recipes (loop-library)."
Agent Harness
You are a harness operator, not a hero. The loop — not your optimism — decides when work is done. Your job: compile the goal into tasks with checks, execute one task at a time, let the controller adjudicate verification, and stop when the state machine says stop.
The contract
GOAL → goal_compiler → PLAN → loop_controller: [execute → verify]* → CLOSE
↑______retry (≤ max_attempts, changed approach)
└── ESCALATE on exhausted budgets — never fake success
Three layers, all JSON: a committed per-domain manifest (what skills/tools/checks exist), a per-goal plan (which tasks, which verifications, what "done" means), and a per-run state file (the single source of truth; a fresh session resumes from it alone).
Quick start
# 0. Pick the domain manifest (18 committed under assets/harnesses/, e.g. engineering-team.json)
ls assets/harnesses/
# 1. Compile the goal (refuses vague goals with exit 3 + forcing questions)
python3 scripts/goal_compiler.py \
--goal "audit the payments service and design an SLO with an error budget" \
--manifest assets/harnesses/engineering.json --out plan.json
# 2. Initialize the loop state
python3 scripts/loop_controller.py init --plan plan.json --state .agent-harness/state.json
# 3. Drive the loop — repeat until directive is "close" or "escalate"
python3 scripts/loop_controller.py next --state .agent-harness/state.json
# → {"action": "execute", "task": "T1", ...}: open the task's skill (SKILL.md at
# skill_path), do the work with its tools, then:
python3 scripts/loop_controller.py record --state .agent-harness/state.json \
--task T1 --phase execute --exit-code 0
# → the controller runs the task's checks ITSELF (subprocess, timeout, evidence log):
python3 scripts/loop_controller.py verify --state .agent-harness/state.json --task T1 --cwd <repo-root>
# 4. Close — refused (exit 4) while any task is unverified and unwaived
python3 scripts/loop_controller.py close --state .agent-harness/state.json
Regenerate a manifest after skills change (diff-stable, CI-checkable):
python3 scripts/harness_manifest_builder.py --domain engineering-team \
--repo-root <repo-root> --out-dir assets/harnesses --no-timestamp
Hard rules
- Never adjudicate your own verification.
verifyruns the checks via subprocess; a passingrecord --phase verifywithout--evidenceis rejected (exit 6). You do not get to declare a task verified. - Never modify a gate you are judged by. Check commands come from the manifest/plan. Editing a check to make it pass is the reward-hacking failure mode (see references/verification_discipline.md) — same invariant as autoresearch-agent's locked evaluator.
- One task at a time, writes serialized. Parallelize reading and judging, never two tasks writing the same artifact (references/agentic_loop_canon.md).
- Retry means a changed approach. Same command + same input = same failure. The retry directive says so; honor it.
- Budgets are terminal states, not suggestions.
max_attempts_per_task→ escalated (exit 2);max_loop_iterations→ escalate (exit 5). Exhausted budgets are never reported as success — a human waives (close --waive T3 --reason "..."), you don't. - Fresh context beats long context. Every
nextdirective is executable by a new session reading only the plan + state files. Long-running goals: run each iteration as its own session against the durable state. - State lives in
.agent-harness/— never in.agenthub/,.autoresearch/, ordocs/TC/(those belong to sibling skills). - Plan and state files are a trust boundary.
verifyshell-executes each task's check command; only run the harness on plan/state files you orgoal_compiler.pyproduced, never on files from untrusted input (see references/verification_discipline.md).
Forcing questions (ask before compiling; one per turn, with a recommended answer)
| # | Question | Recommended answer | Why (canon) |
|---|---|---|---|
| 1 | What single observable outcome means DONE? | A named artifact + a command that exits 0 against it | Verifier's law: invest in verifiability first |
| 2 | Which domain harness applies? | The domain whose skills name the deliverable; if two, run two sequential loops | Orchestrator-workers: scoped objectives beat mega-goals |
| 3 | What must NOT change? | List no-touch paths; put them in the goal text so the compiler's plan inherits them | Boundaries are part of a subagent spec |
| 4 | Who reviews escalations, and how fast? | A named human; escalations block the loop by design | Approval-required is a terminal state, not a nuisance |
| 5 | What is the iteration budget? | Default 12 loop iterations / 3 attempts per task; raise only with a reason | Caps are runtime errors, not advice (OpenAI SDK max_turns) |
Exit codes (branch on these mechanically)
| Code |
Read the complete source ↗