A team that could not previously ship a feature in under three weeks can now produce a working implementation in an afternoon. The interesting part is not the speed. It is that nearly every ritual built around the old cost structure has quietly stopped doing its job.
Estimation assumed effort was the unknown. Ticket decomposition assumed a unit of work had to be small enough for one person to hold. Code review assumed a human read every line before it merged. All three assumptions are now false in different ways, and teams are discovering it in production rather than in planning.
The constraint moved upstream
When execution is cheap, the binding constraint becomes the clarity of the instruction. An agent handed an ambiguous requirement does not slow down and ask a clarifying question in the hallway. It commits to an interpretation and produces a large volume of internally consistent work built on top of it.
The failure mode is not bad code. It is a great deal of good code pointed in the wrong direction.
This is why "just write better prompts" understates the problem. The unit that needs improving is not the prompt, it is the specification that the prompt draws from — and most organisations have never had to write one precisely enough to be machine-actionable.
Three things that have to be written down
Intent: the destination the system is heading toward, written once and checked against constantly. Without it, agents optimise locally and drift globally.
Quality gates: conditions that fire before code is written rather than after. This replaces the review meeting with a contract a machine can check.
Preserved reasoning: decisions captured in-system and versioned, so the next session does not rebuild context from scratch or contradict the last one.
What this changes about the team
The scarce skill becomes judgement about whether output is right, rather than the ability to produce it. That is a different job, and it is not the job most senior engineers were hired into or promoted for. It is closer to editing than to authorship, and it is harder to measure.
We do not think this is settled. The claims above come from our own delivery work and from watching client teams adopt agentic tooling over the past eighteen months. They are the current best version of a moving argument, and we expect parts of it to be wrong.
Codazen Engineering
Written by the team on delivery. This work is published as it happens, with its limits stated — see how work moves.
Published Jul 14, 2026 · revised Aug 2, 2026