How I delegate engineering work to AI agents without losing the plot

The problem

I have a day job. Everything I build independently happens in evenings and weekends, which means the constraint isn’t ideas or skill, it’s hours. Agents change that math, but only if the output is trustworthy enough that reviewing it costs less than writing it.

The operating model

One lead per repository. I run a small CLI, workforce, that gives each project its own persistent Claude Code team lead in a dedicated git worktree, with its own chat channel. TrendVesting and TendForm each get a lead; the lead spins up workers as needed and keeps task notes that survive restarts. My original checkout and uncommitted work are never touched.

Gates, with attribution. In TigerMill, the first rule was “every gate needs a human.” Thirty minutes later I reversed it: agents may close gates, but every review records who acted, and an agent’s approval is labelled as the agent’s, never as a human check. Optimistic concurrency (a revision check that returns a conflict) does the real safety work. Approval becomes an auditable event instead of a bottleneck.

No spend in the toolbelt. TigerMill’s MCP server grew to 13 tools, and none of them can generate an image. Paid generation happens behind a separate, deliberate step. The cheapest way to keep an agent from running up a bill is to not hand it the credit card.

Verified live beats “tests passed.” TendForm has a thin test suite for its size. What actually caught problems was typechecks that block deploys, checking the change on production after it shipped, and having an agent read screenshots of the real UI.

Receipts in the history. Once agents write the commit messages, they can afford to write good ones. TrendVesting’s commit bodies went from essentially empty to a dozen lines of reasoning and production measurements. That history is what let me write the Built with Agents series with commit hashes instead of vibes.

What went wrong

• Delegated taste failed. An agent approved TigerMill art that I then rejected. Agents can close gates on objective checks; subjective acceptance still needs a named human.
• Agent docs rotted. TendForm’s CLAUDE.md kept describing email and infrastructure that were replaced on day one. Instructions are infrastructure and need the same maintenance.
• Allowlists drift toward over-granting. One-off approvals accumulate into policy. They need a periodic review like any other access list.

What it means for a team

The unit that matters isn’t tokens or seats. It’s dollars per accepted output: a merged change, a shipped panel, a resolved ticket, including rerolls and review time. Most teams adopting agents measure none of that. The playbook above is how I’d start.