Harness engineering

Rendered from knowledge/harness-engineering/README.md — the text an agent reads.

Methodology for building systems out of LLM agents and verifying that they work: how to train the artifacts agents read, evaluate agent behavior, structure multi-agent work, and gate what agents ship. Everything here is application-free practice doctrine — how to run each pattern and when it helps. The concrete machinery embodying it in an exobrain — the A/B and self-test skills, the validation gates — is indexed in ../exobrain/machinery.md, and each topic file names its resident implementation; doctrine lives here, mechanics live there. Application layers (production bots, oversight agents, team operating models) are out of scope, and the interview discipline is the meta-domain’s grill.md, not this domain’s.

Start with principles.md — the cross-cutting doctrine everything else embodies. Then:

• eval-loops.md — training any LLM-read artifact against a labeled set: harness validity, splits, ground-truth hygiene, metrics, iteration moves, and deployment follow-through.
• blind-protocols.md — what to hide from which agent, and how to enforce it mechanically.
• adversarial-verification.md — critic/judge structures, validating the instrument itself, negative controls, contained red-teaming.
• behavioral-testing.md — treating agent context as testable code: A/B of context changes, behavioral suites, harness-fidelity probes.
• gates-and-proof.md — layered gates on what agents ship, and the proof bar for anything that reaches beyond its author.
• orchestration.md — structuring multi-agent work: panels and handoffs.
• failure-modes.md — recurring agent failure classes and their engineered counters.