Latest insights on product thinking, engineering excellence, and building software that ships faster and safer.
Your engineering agent reviews PRs, writes tests, and ships code. Nobody can tell you whether it is better than rubber-stamping. Without evals on the agent itself, you are flying blind on the tool that is making 40% of your engineering decisions.
AI code passes review on day one and ages worst by month six — not because it's wrong, but because the design intent that makes it refactorable was never part of the diff. The fix is encoding intent in artifacts the next iteration must read.
The number of days between a new engineer's start date and their first feature in production is the most honest readout of your codebase's health. Every friction point they hit — environment chaos, unclear boundaries, noisy tests, PR limbo — is a structural problem your team learned to route around. The onboarding clock is not running on your new engineer. It's running on your codebase.
When two services integrate, an implicit agreement is formed — a field name, a response shape, an error code. That agreement was never written as a test. Both services keep shipping. The agreement does not update. Contract testing makes the assumption explicit, verifiable, and enforced by CI before a breaking change reaches production.
AI can generate a complete test suite — descriptive names, 94% coverage, green CI — in the same session it writes the implementation. The gap is behavioral: tests written after code verify what the code does, not what it should. Mutation testing makes this visible.
59% of developers ship AI-generated code they don't fully understand. The output looks professional. The reasoning underneath is absent. Here's what broke, why it matters, and how engineering discipline rebuilds the apprenticeship AI bypassed.
A PM and an engineer sit down with four colors of sticky notes. Sixty minutes later they have surfaced every business rule, every edge case, and every unanswered question — and produced acceptance tests an AI agent can build against. No prose. No translation layer. No "the AI guessed wrong" post-mortem.
Most teams deploy agents at stage five speed on stage two or three foundations. AI amplifies whatever it's applied to — strong process gets faster, broken process gets more broken, faster. The question before expanding agent use is not whether. It's where your codebase actually sits.
Cutting software delivery lead time in half does not require a rewrite, a new platform, or a green-field project. Here is the four-step sequence that takes teams from around twelve days to under eight days on the system they already have.