The prompt is dead; long live the execution loop

The unit of work in developer tooling changed this year. Not a question and an answer, but a loop that runs for hours — and the developer's job changed with it.

Coding agents now run execution loops for minutes or hours — planning, editing across repositories, running tests, opening pull requests. The prompt-response cycle carried a hidden tax: the developer had to break work into model-sized pieces and personally carry state between them. That tax is now the agent's to pay.

Duration is what makes it qualitative

An agent that runs for an hour can try something, watch it fail in CI, and revise — the ordinary loop of engineering. The reported 30 to 50 percent reductions in manual coding time come mostly from that self-correction, not from generating text faster. Speed was never the constraint. Recovery was.

Model choice became a runtime decision

Claude Code retook first place partly on per-subagent model control — cheap models for mechanical work, frontier models for hard reasoning. That converts model selection from a purchase into a routing decision, and it is why one competitor holds a benchmark record while sitting second. Harness design now matters more than leaderboard position.

The far end of the same line

And the extreme case is already here: a model running unattended for sixteen days. The competitive axis has shifted from response quality to how long a system can be trusted before someone has to look.

Which quietly redefines the job. Less typist, more editor-in-chief: writing specifications sharp enough to delegate, and reviewing diffs instead of drafting them.

Medium — Dave Patten — The state of AI coding agents (2026) → · Levelop — Best AI coding agents in 2026 →