The job stopped being the prompt. | Thursday, June 25, 2026

The idea is small enough to fit in a sentence, which is exactly why it spread: stop prompting your AI agent by hand, and start designing the loop that prompts it for you. Boris Cherny, who leads Claude Code at Anthropic, put it more bluntly. He does not prompt Claude anymore. He writes the loops that prompt Claude.

This sounds like an inside-baseball developer trend, and on the surface it is. But underneath it is the same story that ran through every other headline this week. Capability is commoditizing. The frontier labs are racing to own the chips, the compute, and the talent precisely because the model itself is becoming the cheap part. Loop engineering is that shift seen from the practitioner's desk: when generating code or text is nearly free, the scarce skill is no longer writing the prompt. It is designing the system that decides what to do next, knows when it is done, and does not set your cloud bill on fire in the process.

A loop, here, is a system that runs an agent in a cycle: act, observe the result, decide the next move, repeat, until a defined goal is met. You set the objective and a stopping condition, and the loop does the iterating. Cherny's point is not that coding got easier. It is that the highest-value work moved. The leverage is no longer in the prompt. It is in the loop.

This is the fourth move in a now-familiar progression. Prompt engineering owned 2022 to 2024. Context engineering, curating everything the model sees at inference, was formalized by Anthropic in late 2025. Harness engineering, building the scaffolding around the agent, took early 2026. Loop engineering is the June 2026 layer, and each step has pushed the human further from the keystroke and closer to the system design.

For enterprise leaders, three consequences matter. First, generation is becoming nearly free, which means engineering value concentrates in judgment: defining the right goal, the right termination condition, the right checks. That changes who you hire and what you reward. Second, the economics invert. A single agent in a loop consumes roughly four times the tokens of a normal chat session, and multi-agent loops up to fifteen times, with parallel loops at scale reaching seven-figure monthly bills. FinOps is no longer downstream of the AI strategy; it is the AI strategy. Third, governance gets harder, because a loop that runs unattended for an hour makes hundreds of decisions no human reviewed in real time.

The pitch for loop engineering is seductive: design a system that runs your agent while you sleep. The failure mode arrived almost as fast as the term. Practitioners are already calling it "loopmaxxing," the assumption that pointing an agent at a problem and letting it iterate indefinitely will eventually produce a correct answer. It is the same mistake as the earlier "tokenmaxxing" phase, when teams believed throwing more inference at a task would resolve hard logic on its own.

It does not, and the reason is precise. A loop needs a binary, checkable exit condition to know when it is finished. "Make the tests pass" works because success is verifiable. "Improve the user experience" or "generate a viral strategy" strips the system of any concrete stopping point, so it never converges and never stops spending. The discipline that separates a reliable loop from a budget incinerator is not the loop itself, it is the guardrails around it: an explicit termination condition, plus secondary limits on cost, iterations, wall-clock time, and stall detection that kills the run when the agent makes the same call three times without progress. The industry is selling loops that run while you sleep. The responsible version is loops that stop on their own before they bankrupt you. Convenience says trust the loop. Durability says bound it.

That is not a coincidence. It is the whole industry telling you, in its own slightly exhausting vocabulary, where the value is going. And it rhymes with everything else that happened this week. OpenAI is building its own chip. A startup committed $6.3 billion to compute through 2029. Two researchers left Google and a trillion-dollar company lost a year of gains in a session. None of those are model announcements. All of them are bets on the layers that surround the model, because the model is becoming the commodity and the advantage is moving to the chips, the compute, the data, the people, and the loops that orchestrate all of it.

So when your engineers come to you talking about loops, do not file it under developer fashion. It is a preview of how the work itself is changing. The leverage in your AI organization is leaving the keystroke and moving into system design, and the budget is going to follow it whether you plan for that or not. The companies that do well in the back half of 2026 will be the ones that noticed the leverage moved, staffed for judgment over generation, and put a cost ceiling on every loop before it ran.