Terms  /  Measurement  /  Recovery Cost
09 · Measurement

Recovery Cost

The time and tokens required to get a session back on track after a failure — State Loss, Execution Hallucination, or Plan Drift.

Recovery Cost measures the penalty for agent failures. When an agent hallucinates that a task is complete, the Recovery Cost includes: the time to discover the hallucination, the tokens spent diagnosing what actually happened, the tokens spent re-doing the work correctly, and the tokens spent getting the agent back on track. Recovery Cost is almost always higher than the original task cost — often 2-5x higher — because recovery requires understanding what went wrong before you can fix it. This makes error prevention far more valuable than error correction.

Example
An agent reports a migration task as complete. Recovery timeline: 30 minutes for the developer to discover it wasn't actually done (they noticed during testing), 15 minutes to figure out what the agent did vs. what it claimed, 20 minutes to roll back the partial work, 45 minutes to have the agent redo the task with closer supervision. Total Recovery Cost: 110 minutes and ~15,000 tokens. The original task would have taken 25 minutes and 4,000 tokens if done correctly the first time. Recovery Cost was 4.4x the original task cost.

More in Measurement