Phase 2: Design

Input: an agent spec (docs/agent-spec.md from agent-scoping). Output: docs/agent-design.md.

Almost every unreliable agent is unreliable for one of three structural reasons: it has too many tools, its context is full of things it does not need, or it was given autonomy it was never evaluated for. This phase is where those get decided.

The governing principle: design the smallest system that could do the jobs in the spec. You can add capability after evals show a gap. You cannot easily remove it after users depend on it.

1. Derive tools from jobs, never the reverse

Take each job from the spec and ask what the agent must be able to do to finish it. That list is your tool surface. Anything not on it does not get exposed, however easy it would be to add.