An agent's shippable capability is set at the review gate, not the generation gate
22% of merged PRs are agent-written, ~0.6% agent-reviewed. That gap is the spec, not a model lag.
Evals
Eval as spec: grade the agent's side effects, not its success story
Judge agents by executable end state and pass^N, not self-reports; the eval harness is the product spec, and verification depth is the constraint.