01
Judgment before topology.
The AI layer is only as good as your model of what quality means in the domain. You cannot prompt-engineer or agent-sprinkle your way around missing ground truth.
Method
How I work
02
The IP is the actionability.
Most teams over-invest in parsing and under-invest in what the system does with what it parsed. Ship the actionability layer earlier than feels comfortable.
03
Split agents by context shape, not by action.
The hardest part of long-horizon agents is not the agents — it is the orchestration layer that decides what each agent needs to know.
04
Raise the ceiling, not the floor.
Mentoring stays compatible with shipping when the mentoring itself produces shippable output. Pair on real production work; let async PR review carry the rest.