# Eval-driven development An engineering loop that converts expected behavior and observed failures into repeatable evaluations combining deterministic checks, environment inspection, security testing, model graders, repeated trials, and human judgment. core-concept · Evaluation & verification · maturing · Reviewed 2026-08-09 ## Definition An engineering loop that converts expected behavior and observed failures into repeatable evaluations combining deterministic checks, environment inspection, security testing, model graders, repeated trials, and human judgment. The test target is the whole agent system: outcome, trajectory, environment, side effects, and policy adherence. The discipline is becoming concrete while measurement standards remain nascent. ## Why it matters The test target is the whole agent system: outcome, trajectory, environment, side effects, and policy adherence. The operating shift is from “A benchmark score” to “A living system-level assurance suite.” ## System anatomy - **Practice 1:** Evaluate outcomes and trajectories - **Practice 2:** Calibrate graders against expert judgment - **Practice 3:** Turn production corrections into regression cases ## Important distinctions - **A benchmark score:** This concept moves the engineering system toward a living system-level assurance suite. ## Implementation signals - Evaluate outcomes and trajectories - Calibrate graders against expert judgment - Turn production corrections into regression cases ## Failure modes - Broken tasks producing authoritative scores - Optimizing the benchmark instead of the operating outcome ## Related knowledge - [Independent verifier systems](https://aisdlc.ai/agentic-engineering/multi-agent-verification) — A separation-of-judgment architecture in which builder agents, evaluator agents, deterministic checks, domain experts, and authorization authorities challenge different failure surfaces. - [Observability & control](https://aisdlc.ai/agentic-engineering/observability-control) — The combined telemetry and enforcement architecture for tracing agent behavior, evaluating policy, obtaining approval, constraining action, revoking authority, quarantining execution, and stopping systems. - [Intent engineering](https://aisdlc.ai/agentic-engineering/intent-engineering) — The practice of turning goals into versioned specifications, constraints, interfaces, invariants, decision rights, acceptance criteria, and testable outcomes before agents implement them. ## Sources and further study - [Anthropic — Demystifying evals for AI agents](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents) — A practical treatment of evaluating trajectories, outcomes, graders, tasks, and agent-environment interaction. **Use:** Official guidance. **Limitation:** This is first-party engineering guidance drawn from Anthropic deployments. Evaluation designs remain task- and environment-specific, and the article does not establish independent verifier ownership by itself. - [NIST — Towards best practices for automated benchmark evaluations](https://www.nist.gov/news-events/news/2026/01/towards-best-practices-automated-benchmark-evaluations) — Guidance on the integrity and limitations of automated benchmark evaluation practices. **Use:** Official guidance. **Limitation:** Guidance describes recommended practice; citation does not prove that a control is implemented or effective in a particular environment. - [OpenAI — Building self-improving tax agents with Codex](https://openai.com/index/building-self-improving-tax-agents-with-codex/) — A feedback-loop pattern that turns practitioner corrections and production traces into evaluations and scoped engineering work. **Use:** First-party case study. **Limitation:** This first-party account documents one organization, product, or implementation context and should not be generalized without local evidence. --- This library synthesizes cited research, standards, official documentation, and clearly attributed practitioner perspectives. Maturity describes the state of a concept—not vendor endorsement, production readiness, or permission to deploy. Benchmarks and demonstrations do not replace use-case evaluation, governed controls, independent verification, or named human release authority.