Module 2 · Building AI Agents: Harness Engineering
Harness Engineering: lessons
Context, permissions, sandboxing, errors and tracing — everything that makes an agent trustworthy outside a demo.
4 lessons · 1h 2m
What this module teaches, what it asks of you first and how you practise it are on its syllabus page.
Open the syllabusContext Management in Agent Loops
18 minTrimming, summarising and offloading when the task runs long — without losing the facts the next step needs.
Included with the track Simulation insidePermissions and Sandboxing
16 minAllowlists, human approval and isolated environments for risky tools — enforced by the harness, not by the model's good behaviour.
Included with the track Simulation insideError Handling and Recovery
14 minError messages that help the model correct itself, retries that do not make things worse, and when to give up.
Included with the track Simulation insideObservability and Tracing
14 minLogging every step, token and cost so failures can be traced — and so you know what a normal run looks like.
Included with the track Simulation inside