Unless you have a governance framework.
We built a feature using a coding agent, then shipped it behind a flag. Tests green. PR merged.
Someone flipped the flag on. It broke! The tests had mocked the exact thing that failed.
The agent did exactly what it was told. Nobody told it enough.
That failure mode pushed my team toward spec-driven development a few months ago. Here’s where that took us, and how to adapt it yourself.
Stage 1: Spec first
No code before product intent (problem, users, non-goals), impact, and acceptance criteria are written down. Nothing ships without it. Pick one feature. Write the spec before touching code. Compare the review cycle to your last five features.
Stage 2: Loop engineering
One agent, one task, one workspace, autonomous and self-correcting against the spec (plan, act, verify, retry) before a human ever sees it. Give one agent one workspace and tie its stop condition to the spec, not to “looks done.”
Stage 3: Graph engineering
Multiple agents handing off through contracts instead of a person routing work between them by hand. Wire two loops together through one typed handoff before you wire ten, and only once stages 1 and 2 hold.
Skip a stage and you don’t move faster.
You just fail faster, with more confidence.
Originally posted on LinkedIn
Keep reading
Why agentic AI fails in enterprise software developmentEssay · 2 min