Correction Ratio
The metric tracking how frequently human intervention is required during agent execution, calculated as interventions divided by total task completions.
Definition
The Correction Ratio tracks how frequently human intervention is required during agent execution. It is calculated as:
Human interventions / Total agent task completions
Every time an operator must step in — whether through a Rescue Mission, a manual code fix, or a spec clarification mid-task — it counts as an intervention. The denominator includes all tasks that reach completion, regardless of whether they required corrections along the way.
The ratio falls into three interpretive ranges:
- Low (below 0.1) — the agent completes more than 90% of tasks without human correction. This indicates effective context engineering, high-quality Live Specs, and a well-tuned Eval Harness.
- Moderate (0.1 to 0.3) — normal for complex or novel work where some ambiguity is unavoidable. Teams should expect this range when agents are working on unfamiliar codebases or tasks with incomplete specifications.
- High (above 0.3) — more than 30% of tasks require human correction, indicating a systemic context issue. Common root causes include stale Golden Samples, ambiguous acceptance criteria in specs, or insufficient architectural documentation in the Context Index.
The most actionable use of the Correction Ratio is tracking it per task type rather than as a single aggregate number. A team might have a low overall ratio but a high ratio specifically for database migration tasks, which would indicate that the context for that task category needs enrichment. This per-type breakdown isolates specific context gaps rather than masking them in an average.
The Correction Ratio complements the Operator Leverage Ratio: a rising Correction Ratio explains why the Operator Leverage Ratio is not improving, because operators are spending their time correcting rather than supervising.