The Harness Is Everything That Survives When the Context Forgets

Short summary

An agent harness is the persistent structure that rebuilds an AI agent's discipline from files every session — standing rules, correction memory, and verification scripts that survive context window resets. The author distinguishes execution-layer harnesses (tool schemas, permissions) from governance layers (rule audits, completion checks) and argues that guides (feedforward rules) and sensors (feedback checks) must not share the same authority to avoid an agent grading its own homework.

• •Agent harnesses persist rules and verification across context resets, unlike prompts which evaporate
• •Governance layer separates guides (rules agent reads) from sensors (scripts that verify outcomes)
• •Rules alone fail without sensors; agents nod at rules then repeat mistakes without enforcement

Generated with AI, which can make mistakes.