Skip to content
Skip to content
Research / Open protocol8 min read

Coding Agent Completion Benchmark: Methodology Before Scores

The seed suite measures false-ready outcomes, escaped defects, repair regressions, human work, latency, and cost. It does not yet justify a leaderboard.

Critique

A coding agent can produce an attractive patch while failing the user’s task. A reviewer can catch a defect and introduce another during repair. A workflow can score well on task success while consuming enough human intervention to erase the advantage. The benchmark therefore evaluates the complete path from public task to final claimed outcome.

The three frozen variants
VariantWhat it isolatesRequired record
Author onlyThe implementation agent’s own completion loopFinal state, checks, usage, time, and human actions
Author + reviewThe effect of a separate review pass without verified repairFindings, false findings, escaped defects, and final state
Verified repairReview, repair, and a new verification passRepair diff, regressions, proof freshness, and final state
Every run must report
  1. 1
    Hidden acceptance success
    Did the final repository state satisfy a private oracle the agent and reviewer could not see?
  2. 2
    Serious escaped defects and false findings
    What consequential defect remained, and what review claim failed human adjudication?
  3. 3
    Repair-induced regressions
    Did the attempt to fix one issue break an adjacent behavior?
  4. 4
    Human intervention
    Record actions and minutes, rather than treating invisible operator rescue as agent success.
  5. 5
    Total cost and completion time
    Include author, reviewer, repair, retries, tool calls, and unavailable cost as null—not zero.

The initial manifest covers omitted requirements, stale evidence after workspace mutation, partial reviewer reports, unavailable behavioral checks, explicit user overrides, and repairs that can regress nearby behavior. These cases validate release invariants. They are not a substitute for diverse repository tasks and repeated model-backed trials.

A benchmark task should expose a public request, a reproducible repository snapshot, hidden acceptance behavior, and a clear rule for what counts as human intervention. Mix straightforward maintenance with changes that require reading neighboring code, preserving compatibility, or handling a failure path. If every task is a toy fixture, the benchmark rewards test pattern matching rather than completion quality.

Reviewers should inspect the final state, the acceptance result, the agent and reviewer records, repair diffs, and operator actions. Blind adjudication helps separate a plausible explanation from a correct outcome. Publish disagreements and exclusions rather than hiding them in a single score.

Publication gate
Freeze repository snapshots and tasksKeep acceptance oracles privateRun paired variants repeatedlyRecord versions, budgets, usage, time, and human actionsBlind-adjudicate defects and false findingsPublish raw records, exclusions, confidence intervals, and failures

Product versions and model identities must be pinned because agent behavior changes quickly. Task order should be randomized. Failed infrastructure runs should be reported separately from product failures. A user override is a valid terminal workflow outcome, but it cannot count as hidden-acceptance success unless the private oracle passes.

Inspect the protocol, then challenge it.
The benchmark contract lives with the code so task design and reporting rules can be reviewed before scores exist.
Independent reviewVerified repairReal repositoriesBuilt for developersLoved by agents