The most expensive thing a coding agent does is not writing a bug. It is ending a turn that sounds finished when it is not: "Deployed!" with no URL checked, three requests with one done, or a list of "next steps for you" that the agent could have done itself.
I run many projects through Claude Code and Codex at the same time, so I built agent-harness: a set of Agent Skills, rules and hooks that change how the agent sizes, finishes and reports work. There is no service behind it; it is files in your agent's config.
The harness reads what you asked and picks how to act:
| You type | The agent acts as | It keeps |
|---|---|---|
| a question or a tiny edit | an engineer | nothing extra |
| "fix the login bug" | engineer + QA | the matching row, if a ledger exists |
| "make it production ready" | the whole product team: spec, build, UX review, QA of every flow, release gates, deploy | READINESS.md |
| "get customers" | the CEO: product, growth, sales, support, money | COMPANY.md |
| "finish all my projects" | a portfolio head with bounded parallel workers | PORTFOLIO.md |
A fix stays a fix. A product request gets a ledger of every user flow and release gate.
Every row in a ledger is verified only with evidence: a test run, phone and desktop screenshots, or a URL check. A Stop hook (done-check.py) reads your last request when the agent tries to stop and sends it back when the work is not verifiably finished.
You write standing approvals once in OWNER.md: money caps, which accounts and channels the agent may use, which decisions it makes alone. Inside them it acts. Outside them it asks one exact thing:
NEEDS YOU (1 of 3): Stripe -> create the account -> unblocks checkout
Moving money, signing and creating new third-party accounts always stay yours.
run-company/scripts/schedule.sh installs a launchd or cron job that runs continue in a project every few hours, with no chat open. Every friction and every owner ask is logged, and a weekly review tests one harness change against the current version on the same task before keeping it.
git clone https://github.com/fainir/agent-harness ~/agent-harness && ~/agent-harness/install.sh
The installer is idempotent and keeps your own content; ./install.sh --uninstall removes everything it added. It works with Claude Code and Codex, and with any agent that reads AGENTS.md and ~/.agents/skills. bash tests/test.sh runs the offline test suite (hooks, learn scripts and a full install and uninstall cycle in a throwaway home).
Limits, honestly: the agent is only as good as the model and the accounts it can reach, and background runs spend model usage while they run. one with schedule.sh --remove <dir>.