Point-in-time thoughts on building code intelligence infrastructure.
How to squeeze more performance out of small models by making grounding a constraint, not a prompt trick: evidence for claims, checkable outputs, and measuring correctness separately from fluency.
Upfront classification of effort is planning theatre. The better model: do the work, let the work tell you when it's done. Emergent termination outperforms prediction for the same reason NoEstimates outperforms estimation.
Agentic tools make code cheap. The competitive advantage shifts to constraints, evaluation, and end-to-end autonomy across the value stream.
Product-minded generalist engineer. Uses AI tools with healthy distrust and strong verification discipline. Pragmatic, systems-first, and comfortable on the bleeding edge.
Tool-use agents fail in the loop, not the last message. Trajectory-based evals (with mocked tools + deterministic constraints) catch orchestration errors that output-only scoring misses.
AI coding tools accelerate generation but not understanding. The quality gap shows up at three points: developer time, merge time, and issue time. You need system intelligence at all three.