Evaluation
How to evaluate consequential AI agents
Completion rate measures whether the agent finished. Consequence correctness measures the world it left behind. These are different numbers and must never be merged.
Published 2 July 2026 · 8 min read
Measure the world, not the transcript
A benchmark that reads the agent's own summary is grading an essay about an outcome. The only defensible signal is the state of the simulated system after the run: balances, digests, permissions, tickets, setpoints.
ConsequenceBench evaluates paired worlds, holding model, prompt, tools, budgets, retries, faults, seed and initial snapshot constant while changing only the execution path.
Four separate scores
Correct consequence: the final source state is safe and policy-correct. Unsafe simulated effects: side effects that should never have happened, never offset by completion. Resolved tasks: fully finished work under the benchmark definition. Tool overhead: the cost of getting there.
Reporting a single blended score hides the trade an operator most needs to see, which is how much completion a governed path gives up to remove unsafe effects.
Safe non-execution is a valid outcome
An agent that stops before a connector call because mandatory proof is missing has produced the correct consequence. Scoring it as a failure trains the whole ecosystem toward optimistic execution.
Publish provenance or publish nothing
Every row should carry a source commit, scenario manifest, model configuration, seed, attempt and failure counts, and a package hash. A number without an inspectable artifact is marketing.