Recalibration 101

Level 1 · Course 8

Your agent drifts. It always will — the current that powers it is the same current that pushes it off course. Five recalibrations in one session is not a bad day. It's data. Every time you catch your agent solving forward, building cases instead of testing, or editing files you didn't ask it to touch — that is a data point about where the harness leaks. The question is whether you collect that data or let it evaporate.

This course installs the recalibration protocol. When you say "recalibrate," your agent stops everything. It reads its drift diagnosis file, names exactly which pattern was active ("I was solving forward — created files instead of answering"), states it to you, and logs it permanently to an oversight log. Not "I need to be more careful." Specific. Named. Written. The log becomes a scar map. But the real power comes later. Once a week — or at whatever pace you choose — you review the log. Are the same three drifts repeating? Is there a through-line? Could a single structural rule solve five recurring incidents? You collect data over time. An afternoon of drifts. A week. A month. You and your agent iterate the harness as you see fit. The protocol doesn't just fix the drift. It builds the machine that prevents the next one.

The term for what you're building is a harness — rules, gates, and patterns that shape the gradient into useful work. Every recalibration tightens the harness. Every drift review finds the next weak point. Over time, your agent gets stronger — not because it drifts less, but because the harness catches it faster and the fixes compound.