AI Harness Engineering

AI harness engineering designs a bounded, testable coding-agent loop with task contracts, repository context, independent controls, and reviewable delivery evidence.

Build agent-navigable repositories with layered instructions, a compact repository map, runnable checks, and durable handoffs that coding agents and reviewers can both trust.

Harden coding agent sandboxes with ephemeral credentials for agents, scoped identity, controlled egress, agent execution isolation, audit evidence, and safe recovery from policy denials.

Scale agent-first software delivery with task claims, isolated CI capacity, review-aware queues, deterministic gates, and measured autonomy.

A practical framework for coding agent evaluation: versioned cases, deterministic graders, multi-trial variance, privacy-aware agent observability, and redaction proof.