What compaction forgets.
When an agent's context fills up, the harness has to make room: drop old turns, hide tool output, or summarise. Canary plants facts in an agent's history, compacts it, and records whether the agent still acts on them.
Four chapters
One argument, in order: make room yourself, watch two real agents after two kinds of compaction, read what the experiments found, then check how it is judged. A seed you pick in one chapter follows you to the next.
What this is, and what it isn't
- What it is
- A flight recorder and canary lab for agent context compaction. It records what is in the window at every step, what compaction dropped, and whether the agent acts differently afterwards.
- What it isn't
- Not a new compaction algorithm, an observability platform, a leaderboard or a framework for building agents; the harness exists only to run test scenarios. The first experiment tried to reproduce published work on our own harness and came out split between models; a preregistered rerun confirmed the effect on the model that showed it. The second compares four ways of making room on the same rule. Every result is reported with its interval, nulls included.
- Where the data comes from
- The hero run, the comparisons and the results are recorded: real model output, replayable from committed recordings. The Lab is simulated: real policy code on real seeded histories, run in your browser, no model; its summaries and its "what real models did" lines are recorded. The policy grid is deterministic: real policy code, no model. The recorder's sample run is a synthetic fixture. Every figure names the file it came from.