Evaluation Feedback Loop

Have an agent generate output, evaluate it with a second LLM call, and loop until quality meets your criteria. Each iteration is checkpointed.

Have an agent generate output, then evaluate it with a second LLM call and loop until the quality meets your criteria. Restate persists each iteration, so if the process crashes, it resumes from the last completed evaluation without re-running earlier iterations.

Example: code generation with quality check

Select your SDK:

A generator agent writes code, then an evaluator agent checks it. If the evaluation fails, the generator retries with the feedback. Each iteration is a durable step.

Each generate and evaluate call is persisted in the journal. If the process crashes after a successful generation but before evaluation, the generated code is replayed from the journal without calling the LLM again.