harness-design

Prevent self-evaluation bias in multi-agent pipelines by separating the Generator from an independent Evaluator calibrated with few-shot scoring. This isolation stops the agent building the feature from validating its own work.

A Sprint Contract defines done criteria, testable success metrics, and scoped iterations to avoid infinite rework in long-running AI applications. It structures handoffs between agents to prevent context degradation across extended development cycles.

Set up Playwright-driven validation by configuring an independent Evaluator workflow to run automated UI and API tests against agent-generated features. This validates testable success criteria through automated acceptance checks during iterative delivery.

Yes, the generator-evaluator pattern applies to full-stack projects requiring iterative feature delivery. It orchestrates a Planner to define intent, a Generator to implement, and an Evaluator to run automated acceptance checks against scoped iterations.

Context degradation happens when long-running agent workflows lose critical state across extended iterations without structured handoff artifacts. Multi-agent harnesses prevent this by enforcing Sprint Contracts that maintain testable success metrics and scoped handoffs.