Last year I made the most consequential engineering decision of my career: we stopped treating AI agents as a productivity tool and started treating them as a unit of organization design.
Most teams bolt an assistant onto their existing process and call it transformation. The gains are real but shallow — maybe 20% on individual tasks. The process itself, built around human handoffs, code review queues, and sprint ceremonies, stays untouched. We went the other way: we restructured four squads around agentic development workflows and rebuilt CI/CD to match.
What actually changed
The unit of work got bigger. When an engineer can delegate implementation to agents, a "task" stops being a function and becomes a feature. Our tickets got fewer and larger. Estimation broke immediately — story points assume human typing speed — and we moved to outcome-based planning instead.
Review moved upstream. LLM-assisted code review catches the mechanical issues before a human ever looks. That freed our senior engineers to review intent — architecture, data contracts, failure modes. The irony: review got more human, not less.
The bottleneck moved to specification. When implementation is cheap, ambiguity is expensive. The engineers who thrived were the ones who could write down what they actually wanted — precisely, with edge cases. Writing became the core engineering skill.
What broke
- Onboarding. Junior engineers learn by doing the small tasks agents now do. We had to design deliberate apprenticeship — pairing juniors with agent-assisted seniors and making them own verification.
- Metrics. PR counts and commit velocity became meaningless overnight. We now measure lead time to a verified outcome, and little else.
- Confidence calibration. Agents produce plausible code at a volume that can outrun your testing culture. We doubled down on contract tests and staging environments before scaling agent usage — in that order, on purpose.
What I'd tell another engineering leader
Target the workflow, not the tool. If your process still assumes every line is hand-typed, an agent is just a faster typist in a slow system. The 10x isn't in the model — it's in redesigning the assembly line around what the model makes cheap.
And keep humans on the part they were always best at: deciding what's worth building.