Harness Engineering

In this topic you will practice Harness Engineering to master the discipline that turns an AI model into a reliable professional agent, wrapping it with the guides and sensors that replace the oscillation between the dazzling and the exasperating with verifiable work of predictable quality.

You will train 10 key competencies:

• Distinguish intelligence (what the vendor trains) from agency (what you build around it) and apply the canonical equation Agent = Model + Harness to understand why a model without a harness delivers polished but hollow work.
• Apply Böckeler's framework by differentiating guides (feedforward controls, before the act) and sensors (feedback controls, after the act) as the two pillars that structure any well-designed external harness.
• Differentiate the internal harness (the vendor's responsibility) from the external one (yours) and operate across the three dimensions of regulation, concentrating effort where 99% of the recoverable value in daily work with agents resides.
• Prepare the environment (pre-flight, accessible materials, bootstrap contract) so that a new agent session passes the cold-start test and identifies project, progress, and next step in under three minutes.
• Write high-signal guides —instruction files, conventions, skills, approved fixtures— that orient the agent before it acts, without falling into either overspecification or saturation of the effective context.
• Design actionable sensors by combining deterministic computational controls and inferential LLM-as-judge controls, so that each verification produces useful reports rather than a binary "pass/fail".
• Operate the steering loop by recording every change in a harness journal: when something fails, iterate on the harness instead of scolding the model, because the model is the only thing you cannot improve.
• Configure permissions, isolation, and operating policies —sandbox, pass-state gating, WIP=1— as a framework that balances the agent's autonomy and human control over what it can touch and when.
• Measure the harness with its own KPIs —rebuild cost, cold-start test, rework rate, drift detected through agentic garbage collection— and recognize the common antipatterns before they erode the system's reliability.
• Decide when NOT to use an agent and precisely delimit the human's role as harness designer, critical reviewer, and sole holder of professional judgment over the final deliverable.