Agent Harness Engineering

"Harness engineering: the surrounding infrastructure determines reliability, not model capability alone." — Faros AI

Same model. Same task. Two different harnesses. Anthropic's own retro-game-maker build ran solo in 20 minutes for $9 — and shipped broken input wiring and dead layout space. The full-harness run took 6 hours and cost $200, over 20x more — and shipped working gameplay, plus a sprite generator and level designer nobody asked for. The model never changed.

"Each new session begins with no memory of what came before — agents need a way to bridge the gap between coding sessions." — Anthropic Engineering

A single continuous session has a hard ceiling: the context window. Without a way to bridge sessions, an agent nearing that ceiling rushes to declare "done" — leaving an untested, undocumented, possibly broken feature behind. Anthropic's fix: an initializer agent sets up a shared feature list and progress file once, then every session after it does one feature, tests it, and leaves a clean handoff for the next.

"Clean state" means "the kind of code that would be appropriate for merging to a main branch: there are no major bugs, the code is orderly and well-documented." — Anthropic Engineering

Without a test-before-exit gate, an agent edits a file, glances at the diff, decides it looks right, and marks the task complete — the same failure pattern whether the model is careful or careless. Anthropic's coding-agent pattern runs a browser-automation test before allowing a feature to be marked done, catching the exact bug a "looks fine" read-through misses.

"A planner breaks the big task into smaller chunks. A generator builds each chunk. An evaluator reviews the output and sends feedback." — Anthropic Engineering

Anthropic's Digital Audio Workstation build ran this three-agent architecture end to end — every phase timed and priced. Click through the real breakdown: a planning pass, then three build/QA rounds, each round's evaluator feedback shaping the next round's build.

"Sprint 3 alone had 27 criteria covering the level editor." — Anthropic Engineering

A continuous, undecomposed generation run gave Anthropic's team no checkpoint to catch quality drift until the whole run finished — one comparison ran over two hours before any evaluation happened. Breaking the same work into sprints, each gated by its own evaluator criteria, surfaces problems chunk by chunk instead of at the end. Chunking is a harness decision — the model never "decided" to checkpoint itself.

"This memory is 13 days old — verify against current code before asserting as fact." — a real guardrail from this author's own Claude Code memory system

Harness engineering isn't unique to Anthropic's published case studies — it's in the repo this exact chapter was built in. A CLAUDE.md file with a PII guardrail. A memory system that flags its own staleness before you trust it. A custom skill that encodes an entire build process so it doesn't have to be re-explained every time. Step through a real (paraphrased) session opening this project.