The Long Horizon

Compaction. When the conversation approaches the limit, summarise it and restart the window pre-loaded with the summary. Keep architectural decisions, unresolved bugs, live implementation state; drop resolved tool calls and raw results. Right when the value is in the conversation — extended back-and-forth where continuity of dialogue is the thing.

Structured note-taking. The agent writes its own state to a file outside the window — a NOTES.md, a checklist — and reads it back after a reset. Right when the value is in milestones: long iterative work with defined checkpoints, where the dialogue matters less than knowing what is done. This is how an agent playing Pokémon kept maps and training targets across thousands of steps and multiple resets.

Sub-agents. A lead agent holds the plan; specialists do tool-heavy exploration in their own clean windows and return a condensed summary of one or two thousand tokens. Right for multi-hop research, where parallel exploration pays and the lead's window must stay clean.

The three are not ranked. They match different shapes of work, and picking by shape rather than by fashion is most of the skill.

Compaction has a prescribed order, and it is not the intuitive one.

Maximise recall first: make sure nothing load-bearing is lost, even if the summary is bloated. Only then tighten for precision, stripping what turned out to be superfluous.

The reason is asymmetry of consequence. A summary that keeps too much costs tokens, which is recoverable on the next pass. A summary that dropped the unresolved bug has destroyed information that no longer exists anywhere, and the agent will rediscover it expensively or not at all.

The safest thing to cut, and the first: resolved tool calls and their raw results. Once an action is done and its outcome absorbed, the transcript of it is pure weight.