Agent Harness — Development

Agent Harness turns any domain folder of skills into a bounded agentic loop: it compiles a goal into a verifiable task plan, executes tasks with the domain's own tools, verifies each task with machine-run checks, retries with caps, escalates to a human when budgets exhaust, and refuses to close until everything is verified or explicitly waived. It produces a durable state file and a verified close or an escalation.

engineering claude-code

📂 View skill on GitHub ↗ 🚀 How to use

✨ What it does

Compiles goals into task plans with verifiable checks (refuses vague goals).
Executes tasks one at a time with domain-specific skills and tools.
Runs verification checks via subprocess with timeouts and evidence logging.
Enforces retry caps and iteration budgets, escalating to a human when exhausted.
Refuses to close until all tasks are verified or explicitly waived.
Maintains a durable state file for resuming across sessions.

🎯 When to use it

When you want an agent or subagent to pick up a goal and drive it to a verified close across one of the repo's 18 domains.
When you need a self-verifying loop that enforces machine-run checks and retry caps.
When you want to run a goal through the engineering harness or set up an agentic loop for a specific domain.
When you need to make a domain folder self-verifying with a committed manifest.

🚀 How to use

Trigger by asking to run a goal through the harness, e.g., 'run this goal through the engineering harness' or 'set up an agentic loop for marketing work'. Provide a clear goal, choose a domain manifest, and optionally answer forcing questions. Use the provided Python scripts: goal_compiler.py to compile the goal, loop_controller.py to initialize and drive the loop, and record/verify/close commands. Example prompts:

Run the goal 'audit the payments service and design an SLO with an error budget' through the engineering harness.
Set up an agentic loop for the finance domain to reconcile accounts.

📄 Output: A verified task plan and state file, leading to a close or escalation report.

📦 Add this skill to Claude Code

# 1. Get the skills repo
git clone --depth 1 https://github.com/alirezarezvani/claude-skills /tmp/claude-skills

# 2. Copy this skill into your project (or ~/.claude/skills for all projects)
mkdir -p .claude/skills
cp -r /tmp/claude-skills/engineering/agent-harness/skills/agent-harness .claude/skills/agent-harness

Skill source: engineering/agent-harness/skills/agent-harness/SKILL.md

⚠️ Good to know

Requires a committed domain manifest and the Python scripts; only run on plan/state files you or goal_compiler.py produced, never untrusted input.

❓ FAQ

What happens if a task fails verification repeatedly?

The loop retries with a changed approach up to max_attempts_per_task; if exhausted, it escalates to a human (exit 2).

Can I close the loop with unverified tasks?

No, close is refused (exit 4) while any task is unverified and unwaived; a human must explicitly waive a task.

How does the harness ensure verification is not faked?

The verify command runs checks via subprocess, and recording a verify without evidence is rejected (exit 6).

🤖 Overview, features, install steps and FAQ were generated from the project's SKILL.md on Sep 4, 2026. Always check the original source before running commands.