Harness & Loop Engineering
Prompt engineering asks what to say to the model. Context engineering asks what the model should see. Harness and loop engineering ask a different question: what system should run around the model so it stops making the same mistake twice, and keeps working toward a goal without you typing the next instruction by hand. This is where AI engineering effort has moved in 2026 - not better prompts, but better infrastructure around the model.
The maturity ladder
Prompt engineering (what to say) → context engineering (what the model sees) → harness engineering (the system that catches its mistakes) → loop engineering (the control cycle that runs it autonomously). Each layer doesn't replace the one before - it sits on top of it.
In This Section
Harness Engineering
The five-layer system around a model - tool orchestration, verification, memory, guardrails, observability - that turns one-off fixes into permanent behaviour changes.
Loop Engineering
A task plus a check, repeated. How to design the trigger-act-verify-retry cycle that replaces turn-by-turn prompting.
Official Agent SDKs
Claude Agent SDK and OpenAI AgentKit - first-party, batteries-included harnesses from the model providers themselves.