Managing agents is managing engineers

The first time I handed a non-trivial ticket to a Claude Code agent and walked away, it finished the work, opened the PR, and moved on. No updates, no blockers, no Slack pings. I came back expecting to review code; I ended up reviewing my own management habits.

Everything that makes a senior engineer effective makes an agent effective: a narrow, well-scoped task; acceptance criteria written before work starts; access to the same tools a human would use; a feedback loop. The things that break engineers break agents harder — fuzzy requirements, missing context, permission friction, no way to verify success.

The failure modes rhyme too. Agents hallucinate the same way juniors over-commit in sprint planning: confident, specific, wrong. The fix is the same — make them show their work. Ask for the plan before the code, the test before the fix, the rollback before the deploy. Pair the agent with a linter and a test suite and you get a surprisingly disciplined teammate. Give it nothing to check against and you get a confident mess.

The leverage changes too. A human engineer's throughput is bounded by hours in a day. An agent's is bounded by clarity of instruction. If you can write a clear brief, you can run ten of them in parallel. If you can't, you'll run one badly ten times. The skill that matters in this new stack isn't prompt engineering — it's exactly the same muscle as writing good tickets, which we mostly haven't gotten around to learning.

So the thing engineering managers have been doing badly for twenty years is suddenly load-bearing. Good news: the playbook is the one we already have. Bad news: the people who were getting by on charisma and hope now have a boss that doesn't respond to charisma.