An annoying AI failure mode is drift in long-running sessions. You know how a 3-hour meeting makes your mind go crazy? The same thing happens to the agent.
In long sessions, AI seems to optimize for “finish fast” instead of “follow intent”:
- stops early and says “done”
- modifies tests to make green checks
- takes implementation actions while still in planning mode
- ignores constraints given in instructions
This is drift — a symptom of context pressure. Treat it as a signal, not a surprise.
A practical rule: when you see one of these, restart with a tighter task boundary. You haven’t lost your understanding — you can feed that into the next session. The agent is the one that needs the fresh start.
And move critical rules from chat instructions to enforceable checks:
- run tests + typecheck before done
- block risky actions with permissions
- review diffs in tests and config files before merge
Don’t expect instructions to be followed perfectly forever. Assume drift, then design guardrails.
💡 Try this: Add a “drift trigger” to your workflow: if the agent edits existing test assertions or claims done without verification output, restart in a fresh session.
🔗 Copilot CLI — manage permissions 🔗 Copilot CLI — plan and autopilot modes
How do you handle drift in long sessions?
🟢 Restart fast when behavior slips
🟡 Keep correcting in the same session
🔴 Usually finish manually
⚪ Haven’t noticed this much (perhaps because I keep sessions short?)
What drift symptom do you see most often?