# Review the Plan Not Just the Output ## Claim Adversarial review belongs at the earliest inspectable artifact — the plan — rather than only at the generated output, because the cost of a defect rises with how much work has been built on top of it before anyone catches it. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds where the plan is genuinely inspectable and specific enough to contain checkable claims — a vague plan has nothing concrete for a reviewer to catch. - Breaks where the reviewing agent shares the generating agent's blind spots (same model, same context, same training biases) — in that case plan review ratifies rather than catches, producing false confidence rather than a real check. - Breaks for exploratory work where the plan is expected to be wrong and revised through building, not through pre-review — plan review presumes the plan is meant to be executed close to as-written. - Does not replace output review; it moves the first and cheapest check earlier in the pipeline, it doesn't make later review unnecessary. ## Evidence Joey Kudish describes running multiple review agents over generated plans — checking for defects, inconsistencies, and discrepancies, looping until clear — before any orchestrator spawns implementation agents. This inverts the more common practice of applying review only to generated code, after the cost of a defect has already been paid in wasted implementation effort. > [!quote] > I have multiple review agents look over generated plans for defects, inconsistencies, discrepancies ## Links - [[3. Resources/Permanent Notes/Constraints Validated At Definition Time Propagate Not Recur]] — because extends; that note is the general "shift validation earlier" principle from a design-tokens context, this note is the specific application to agentic-development plan review. - [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]] — because applies-in; plan-stage review only works if the reviewing agent is genuinely independent of the generating agent's context, the same independence condition that note establishes for output review. - [[3. Resources/Permanent Notes/Gate Contributions on Demonstrated Comprehension Not on Tooling]] — because extends; both gate work on a cheap, inspectable comprehension artifact produced before the expensive part happens — a plan here, a one-paragraph summary there. - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — because applies-in. ## Tests / falsifiers - Controlled experiment with seeded plan defects: compare defects caught and token cost between review-at-plan and review-at-code arms, blinding reviewers to which arm they're in — if review-at-plan doesn't catch more defects per unit cost, the claim's practical value is overstated. - Check whether a reviewing agent sharing the generating agent's model and context actually catches fewer defects than an independent one — if plan review performs similarly regardless of reviewer independence, the "shares blind spots" breaking condition above may not matter as much as stated. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Literature Notes/Joey Kudish - how I use AI these days]] — Kudish's description of running review agents over generated plans before implementation.