Managing AI coding agents is becoming a real job

The loop nobody named

Track what you actually do all day when you work with AI coding agents, and a pattern falls out that has nothing to do with typing code. I run it dozens of times a day now:

1. Brief. Turn a fuzzy goal into an outcome with constraints: what done means, what must not change, where the edges are.
2. Delegate. Pick the right agent for the job, hand over the task, resist the urge to micro-instruct the how.
3. Inspect. Watch the trajectory, skim the diffs, test the behavior. Push back on scope creep. Reject work that merely looks finished.
4. Unblock. Answer the question the agent is stuck on, approve the risky command, make the design decision it cannot make for you.

Brief, delegate, inspect, unblock. Any engineering manager will recognize the shape, because it is their job description with the humans swapped out. The hands-on-keyboard hours shrank; the judgment hours grew. The role under the role changed before anyone updated the titles.

What transfers (it is not prompt tricks)

The developers I watch thrive with agents are not the ones with a folder of magic prompts. Prompt phrasing matters far less than three older skills.

Writing specifications. The brief is the product now. People who can state acceptance criteria, name the non-goals, and anticipate the ambiguous cases get radically better output from the same models. This was always valuable; agents turned it into the highest-leverage hour of the day.

Code review instincts. Reading a diff for what is missing, noticing the test that got easier, smelling the dependency that did not need to exist. Review used to be a chore between coding sessions. It is now the core verification act, and the people who were good at it are suddenly the safest pair of hands in the room.

Operational thinking. Budgets, blast radius, rollback plans. Agents introduce a live cost meter and a real possibility of compounding mistakes, so the sre mindset of "limit damage, keep receipts" maps directly onto daily coding for the first time.

The new literacies

Two genuinely new skills sit on top. The first is knowing your agents' temperaments. Claude Code, Codex CLI and Gemini CLI fail differently, plan differently, and deserve different kinds of tasks; running several of them teaches you a roster the way a manager learns a team. The second is parallelism judgment: knowing when five agents on five lanes is a gift and when it is five times the mess. I wrote about the mechanics of that in the multi-agent field guide.

Managers get dashboards. We got scrollback.

Here is the gap that bothers me. Every real management role comes with instruments: standups, boards, review queues, budgets. The agent manager role mostly comes with terminal scrollback. You are asked to supervise a team you cannot see, staff it with workers who never say no, and account for spend you learn about at the end of the month.

That gap is the reason The Termi Protocol exists. Rooms instead of tabs, task boards instead of mental queues, checkpoints instead of hope, live cost instead of invoice surprises. If the job is becoming management, the job deserves management tools, preferably ones that are actually pleasant to look at for eight hours.