Eval-Driven Development

Definition

Eval-driven development makes representative evaluation data, explicit criteria, and release evidence part of the product-development loop. Changes to models, prompts, tools, or workflows are compared against task-specific expectations before release and monitored again in production. Harvey ties evals directly to launch decisions, while Sword starts with datasets and human criteria, validates automated evaluators, gates releases, and reuses the same machinery online. Harvey Sword Health

Corpus Support

The analysis finds 475 deduplicated agentic systems (47.9%) across 16 industries, supported by 291 article, 404 transcript, and 85 summary evidence items; 775 are labeled stated and 5 implied. Corpus analysis

Problem It Solves

LLM behavior changes nondeterministically and generic benchmarks rarely represent a product’s actual tasks. Custom evals create a repeatable basis for deciding whether a change improves the system, regresses an edge case, or is ready to ship. Parlance rejects leaderboards and vibe checks as substitutes for project-specific evaluation; Sierra repeatedly simulates conversations because one passing run cannot characterize an agent. Parlance Labs Sierra

Applicability Signals

• Model, prompt, retrieval, or tool changes can alter user-visible behavior without changing application code. Harvey
• The task is nondeterministic enough that pass-once unit tests hide outcome variance. Sierra
• Domain experts can articulate quality criteria that generic benchmarks miss. Sword Health
• Production failures should become durable regression cases. Deutsche Telekom

Architecture and Variants

• Offline release suite: curated examples, deterministic assertions, model judges, and human labels gate model or prompt changes. Harvey
• Repeated simulation: run the same conversational task several times, including adversarial and channel-specific cases. Sierra
• Component plus end-to-end evals: evaluate query rewriting, retrieval, generation, and citations separately, then retain workflow-level tests. Harvey
• Online feedback loop: log production traces, retrieved documents, user feedback, and evaluator results so failures can be promoted into the next dataset. Sword Health

Why Teams Choose It

The pattern converts subjective iteration into a release decision process while preserving human judgment for dimensions that automated metrics do not capture. It also makes model/provider changes comparable on the product’s actual workload rather than on unrelated benchmark averages. Parlance Labs Harvey

Tradeoffs

• Human evaluation provides domain fidelity but is expensive and slow; Harvey decomposes workflows and uses automated judges to reserve expert attention for the hardest decisions. Harvey
• Automated evaluators scale, but they can optimize the wrong criterion unless calibrated against expert labels. Sword Health
• Exhaustively evaluating every intermediate step creates brittle tests; Parlance favors final interactions unless a component boundary has a clear independent contract. Parlance Labs

Gotchas and Resolutions

• Randomly sampled datasets hide rare healthcare failures; Sword deliberately selects hard and boundary cases. Sword Health
• An evaluator suite can pass while overall product quality still feels wrong, so major releases retain pairwise review and a human “vibe check.” Sword Health
• Flagged production questions disappear unless they become test cases; Deutsche Telekom feeds unresolved interactions into future release suites. Deutsche Telekom

When Not to Use It

Do not create a heavy model-judge stack before the team can define representative tasks and failure criteria. Early assertions can become useless as the product changes, and automated grading can become another form of vibe checking if it is not calibrated. Start with inspectable examples and human decisions, then automate stable criteria. Parlance Labs

Boundaries and Related Patterns

llm-as-judge is one evaluator implementation; eval-driven development is the broader lifecycle that owns datasets, criteria, release decisions, and feedback. observability-tracing records what happened, while evaluation decides whether it was acceptable. data-quality-loop improves the examples or labels feeding models; it may be driven by eval findings but is not the release discipline itself. Sword Health

Operational Checklist

• Define task-specific examples and acceptance criteria before comparing models. Parlance Labs
• Record evaluator-to-human agreement, not only evaluator scores. Sword Health
• Run repeated trials for nondeterministic and conversational paths. Sierra
• Promote production failures and unresolved interactions into regression datasets. Deutsche Telekom
• Keep an explicit human release decision for high-impact domains. Harvey

Representative Implementations

Methodology and Limitations

This synthesis uses a small cross-industry sample of the 475 supporting systems. Evaluation practices are often reported by teams with mature quality programs, so frequency and sophistication may be publication-biased. The page describes recurring architecture and process choices; it does not establish that any evaluator is objectively valid outside its cited domain. Corpus methodology