Discover Broadly, Implement Narrowly

Short summary

This post proposes a governance framework for AI coding agents: separate observation authority (which should be wide) from action authority (which should be narrow). Agents should surface architectural problems they discover during implementation—duplicated logic, missing boundaries, lifecycle drift—but leave architectural decisions to humans. This shifts human burden from reconstructing all generated code to adjudicating grounded architectural evidence.

• •Separate observation authority (broad) from action authority (narrow) when using coding agents
• •Agents should surface architectural findings but humans decide on changes
• •Define architectural evidence as implementation data that changes confidence in architectural propositions

Generated with AI, which can make mistakes.