技能 编程开发 智能体目标执行框架

智能体目标执行框架

v20260717
agent-harness
智能体目标执行框架提供了一个强大的可验证循环结构,用于推动跨越专业领域的复杂目标。它将高层目标转换为可验证的任务计划,使用领域特定的工具按顺序执行任务,并通过机器运行检查严格验证每一步。系统包含重试、预算管理和强制人工升级机制,确保目标只有在所有进展都可验证完成或明确豁免后才能关闭。
获取技能
480 次下载
概览

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

  1. Never adjudicate your own verification. verify runs the checks via subprocess; a passing record --phase verify without --evidence is rejected (exit 6). You do not get to declare a task verified.
  2. 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.
  3. One task at a time, writes serialized. Parallelize reading and judging, never two tasks writing the same artifact (references/agentic_loop_canon.md).
  4. Retry means a changed approach. Same command + same input = same failure. The retry directive says so; honor it.
  5. 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.
  6. Fresh context beats long context. Every next directive 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.
  7. State lives in .agent-harness/ — never in .agenthub/, .autoresearch/, or docs/TC/ (those belong to sibling skills).
  8. Plan and state files are a trust boundary. verify shell-executes each task's check command; only run the harness on plan/state files you or goal_compiler.py produced, 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 Tool Meaning
0 all OK / directive emitted
2 loop_controller Escalation required — a human must review the evidence log
3 goal_compiler Goal too vague — answer the forcing questions, recompile
4 goal_compiler / loop_controller No skill matched / close refused (unverified tasks)
5 loop_controller Global iteration cap reached
6 loop_controller Invalid transition (recording on verified task, evidence missing, unknown task)

Verifiable success

  • python3 scripts/harness_manifest_builder.py --sample, scripts/goal_compiler.py --sample, and scripts/loop_controller.py --sample all exit 0.
  • A vague goal (--goal "make it better") exits 3 and prints forcing questions.
  • loop_controller.py close on a state with an unverified task exits 4.
  • The demo loop in loop_controller.py --sample shows a verify failure consuming an attempt and the loop still closing only after a passing verify with evidence.

Related skills

  • workflow-builder: authoring deterministic .js scripts for Claude Code's Workflow tool. NOT for goal-to-close loop state (this skill).
  • agenthub: N parallel agents competing on ONE task in git worktrees. Use it inside a harness task that wants competing attempts.
  • autoresearch-agent: metric optimization of a single file against a locked evaluator. Use it when a task's done_when is "metric improves".
  • tc-tracker: per-code-change lifecycle records. Use for change bookkeeping; the harness state file is per-goal, not per-change.
  • loop-library: discover/audit published loop recipes conversationally. This skill is the executable enforcement of that vocabulary.
  • ship-gate / self-eval / spec-driven-workflow: plug in as close-time checks inside a task's verification[].

See references/domain_harness_design.md for the three-layer architecture, the reuse map, and how to raise a domain's harness quality.

信息
Category 编程开发
Name agent-harness
版本 v20260717
大小 119.15KB
更新时间 2026-07-18
语言