What I mean by "agentic engineer"

Mon, 21 Sep 2026 00:00:00 +0000

Every team I’ve talked to this year is running agents. Almost none of them can tell you whether last week’s change made things better or worse. That gap is the whole job.

The loop, not the model

Model choice is the least interesting variable. What separates a setup that works from one that flails is whether the agent can find out it was wrong:

// fail loudly, and early
async function runLoop(task: Task, maxTurns = 12) {
 for (let turn = 0; turn < maxTurns; turn++) {
 const r = await agent.step(task)
 if (r.testsPassed) return r // done
 task = task.withFeedback(r.stderr)
 }
 throw new Error("no convergence")
}

A model without a feedback loop is a very confident random number generator.

The useful skills turn out to be boring ones:

Writing CLAUDE.md files that survive contact with reality
Making failures cheap and legible
Knowing when to stop the loop and read the code yourself

None of that is new. It is the same discipline as any other engineering work, with a much faster and much less reliable collaborator.

carraes.dev

What I mean by "agentic engineer"

The loop, not the model