- ai
- workflow
Notes on agentic development workflows
How I delegate well-scoped work to AI agents — small, verifiable slices, tight feedback loops, and tests as the contract.
Small, verifiable slices
The unit of agentic work that actually ships is a thin vertical slice: one behavior, end to end, with a test that proves it. Big ambiguous tasks stall; narrow ones with a clear definition of done move fast.
Tests as the contract
When an agent and I agree on a test first, the test becomes the spec. The agent has an objective target, and I have a cheap way to verify the result without reading every line.
Tight feedback loops
The faster the loop — type-check, lint, test — the more autonomy an agent can safely take. Most of the value is in making verification cheap, not in writing more prompts.
Keep humans on the seams
I review at the boundaries: schemas, public APIs, the edges where a mistake propagates. Implementation details inside a well-tested seam matter less than the shape of the seam itself.