Many AI agent failures come from acting on incomplete info, not bad reasoning
Many failures blamed on ' reasoning aren't reasoning problems at all — they happen because the agent executes an action while key information is still missing. Rather than proposing a new model or , this describes a lightweight execution pattern built on four steps: separate state (data) from execution logic; never let the AI guess at missing information — mark it explicitly as "Unknown" instead; block execution if even one Unknown remains; and let the final state itself serve as the execution record, functioning as an .
Under this pattern, the AI's job shifts from inferring gaps to matching against already-confirmed information. If something is unknown, the user fills it in — not the model.
No new , , or language is needed; a plain JSON structure is enough. An example shows a login-bug-fix task represented as JSON tracking whether the modification scope is defined, whether test criteria exist, and user like "do not change UI."
Key points
- Agent failures often stem from executing with incomplete inputs, not from weak reasoning
- Proposes a Separation → → Enforcement → Traceability pattern
- Missing info must be marked explicitly as "Unknown" rather than guessed; any remaining Unknown blocks execution
- The user, not the AI, is responsible for filling in unknowns
- Needs no new model, , or — a plain JSON structure suffices