Amey-Thakur Agent session management
Long sessions accumulate context, drift from the original goal, and carry earlier mistakes forward. Managing them means deciding when to continue, when to summarise, and when to start again.
Method
- Start a new session per distinct task. Carrying an unrelated history costs tokens and biases the agent toward the previous problem.
- Commit at every working state. Frequent commits make it safe to discard a session that went wrong (see git-workflow).
- Summarise and restart when context fills. A fresh session with a written summary of decisions beats a long one where early context has been squeezed out (see conversation-memory).
- Restate the goal periodically. Long sessions drift, and a mid-session restatement is cheap insurance (see goal-driven-execution).
- Externalise state into files. A plan or notes file in the repository survives the session and is readable by the next one (see agent-blackboard-workspace).
- Stop when the agent is looping. Repeated failed attempts on the same problem will not resolve with another turn, and the fix is more context or a different approach.
- Review before accepting a long session's work. The cumulative diff may contain changes made early and forgotten (see agent-refactoring-workflow).
Boundaries
Session management mitigates context limits and does not remove them. Restarting loses implicit understanding built during the session, which is why externalised notes matter. Tools differ in how they handle context compaction.