Agent evaluation
Repeatable testing of an AI agent's complete task, tool use, permissions and final system state against explicit acceptance criteria. A successful demo is not a production reliability estimate.
- Evidence
- Representative tasks, traces, final-state checks, failure categories
- Release boundary
- Defined scope, critical failure rules, monitored pilot
Use deterministic outcome checks when possible, calibrated model graders where useful, and human review where meaning or consequences require it. Version the model, prompt, tools and environment.
Averages can hide critical failures. Keep authority violations and dangerous actions separate from ordinary task success, and rerun the suite when the system changes.
Connected
- Claude CodeAnthropic's agentic coding tool that reads a codebase, edits files, runs commands and verifies its work from the terminal, IDE or web.
- AI task time horizonMETR's measure of the length of task, in human working time, that a model completes with 50% reliability; about one hour for frontier models in March 2025, doubling roughly every seven months.
- AI data retentionThe rules governing which AI inputs, outputs, logs and derived records remain in a system, for what purpose, and for how long. Not using data for training does not mean that it is not stored.
- Accessibility testingEvaluation of digital accessibility through automated checks, knowledgeable manual assessment and research with people with different access needs. Conformance evaluation and usability research provide complementary evidence.
Linked from
- Outcome-based promptingAsking an AI agent to pursue a specified change for a user or system, with evidence and boundaries, instead of merely producing a named deliverable.
- AI outside judgeAn acceptance check outside an agent’s self-assessment, such as a deterministic test, held-out evaluation, measurement or separately briefed reviewer.
- Prompt injectionAn attempt to make an AI system follow instructions embedded in untrusted material such as email, documents or websites instead of the authorized user’s task.
Field notes that cover Agent evaluation
More concepts & methods
- machine learningscientific study of algorithms and statistical models that computer systems use to perform tasks without explicit instructions
- deep learningspecialized form of machine learning that uses artificial neural networks to identify patterns in data
- genetic algorithmcompetitive algorithm for searching a problem space
- Altitude, depth and rangeFelix Rieseberg’s framework for ambition with AI models: speed repeats work faster, altitude targets outcomes, depth uses expert instruments and range combines disciplines.
- Multidisciplinary agent reviewA workflow in which specialist passes apply distinct methods and produce concrete evidence, rather than merely imitate expert personas.
- Zero data retentionA provider arrangement limiting storage of eligible AI request content, subject to approval, feature coverage and documented exceptions. It does not remove records retained by the customer's own application.