AINEWS
The Salty Lesson: Stop Prompting, Start Designing Loops
Three senior engineers say the next leverage isn't a better prompt, it's a loop that prompts for you. The harder question is which loop to climb.
Three senior engineers say the next leverage isn't a better prompt, it's a loop that prompts for you. The harder question is which loop to climb.
Engineers building with AI agents are running out of the one resource they keep trying to spend: themselves. The new leverage is not a better prompt. It is a loop that does the prompting for you.
That is the throughline in a recent Latent Space AINews dispatch on Loopcraft, which gathers three short public statements into a single working thesis. Anthropic's Claude Code lead Boris Cherny wrote that he no longer prompts Claude. "I write loops," he said. "The loops do the work." Peter Steinberger, a long-time developer-tools operator, sharpened the reframe: "You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents." Andrej Karpathy, in a long-form YouTube discussion on code agents and AutoResearch, put the same idea in operational terms. The goal, he argued, is to "remove yourself as the bottleneck," "maximize your token throughput and not be in the loop," and "arrange it once and hit go." The human steps out of the inner loop and into the outer loop, where the work is to design the system that runs the work.
The newsletter wraps this with a coinage worth taking seriously on its own terms: a "Salty Lesson for agents," a riff on Rich Sutton's Bitter Lesson. The Salty Lesson, as the Latent Space editor frames it, is to stop fixing things yourself and instead build systems that scale with more agents: goals, orchestration, harness, the scaffolding that compounds. Hand-fixing is a trap because it does not compound. Loops compound.
The interesting second-order question is not whether loops are better than prompts. They clearly are, for some definition of better, in the cases Cherny, Steinberger, and Karpathy are describing. The question is which loops, and at what altitude. A practitioner is always sitting inside a stack of loops already: continuous integration, eval harnesses, repo hygiene, sandboxing, model routing, retry and backoff. The new skill is choosing where to invest the next unit of engineering. Go up a loop, and you buy leverage. Your goal-and-orchestration layer subsumes dozens of hand-tuned sub-loops. Go down a loop, and you buy reliability. You hand-tune a sub-component, a grader, a prompt template, a guardrail, because predictability matters more than velocity at that altitude.
The two moves pull in opposite directions. Going up is a bet that model capability will keep improving, so the cost of a thin orchestration layer keeps falling. Going down is a bet that this particular agent, on this particular task, on this particular day, will behave like the version you measured, so the cost of a fat reliability layer is worth paying. A team that always goes up ships features but ships surprises. A team that always goes down ships certainty but ships slowly. The Salty Lesson is a reminder that going up, deliberately, is the move that scales with more agents. It is not a license to skip the going-down work that the source itself flags. As the same Latent Space issue concedes, the field is "already in" the messy part: nested loops that fail in nested ways.
The Latent Space framing is editorial, not consensus. It is a curated read of three short public posts, with the "Salty Lesson" name attached by the newsletter, not by Sutton or by any of the people quoted. That is a fair piece of synthesis to build on, but a reader should know that the strongest claim in the piece, that hand-fixing is a trap, is the editor's argument, not a verified finding. It also has a recruitment edge: telling a working engineer to stop doing the work they are good at, in favor of work they are not yet good at, is a way to move a category. The earlier Latent Space Autoresearch: Sparks of Recursive Self Improvement issue, from March 2026, was the same move at a different altitude, framing agentic research loops as the next AutoML moment.
The watch item is concrete. Look for shipped systems, not more posts, that put a human at the outer loop and let the inner loop fail and recover on its own. Look for benchmarks that measure the cost of the outer loop itself, the time the human spends designing and revising the loop. Look for failure modes that come specifically from going up a loop too eagerly: silent degradation, eval gaming, harness lock-in. The Salty Lesson will earn its keep when the harness engineers, not the prompt engineers, are the most expensive people on the team.