Overview
Agentic products hit the same walls: flaky tool calls, failures nobody can see, no evals, no safe way to ship a prompt change. A harness around the agent makes it behave like a production system: a typed tool layer, traces you can actually read, and an eval suite that catches regressions before users do.
What this covers
- Typed, versioned tool layer (testable in isolation)
- An example keyword-search tool (
kb_search) that demonstrates the eval harness on a retrieval-style task - One structured trace line per tool call, small enough to grep
- Regression eval suite wired into CI
Status: v0.1, deliberately minimal. A reference implementation of the pattern, plain enough to copy into your own codebase rather than depend on.
Where it stands
agent-harness is open source and deliberately minimal: a typed tool layer and a regression suite, not a framework.