← All writing

Stopping a context optimization

A representative compaction test lost important state, so I stopped the experiment and returned to shorter work stages.

I wanted to reduce the amount of context my agents carried through a session. The idea was to preserve the important facts, compact the rest, and give the agent enough information to finish the current task.

The implementation started getting more elaborate than the problem justified. I asked to keep the tests tightly scoped and check the idea against the way I actually work: an orchestrator handing bounded tasks to subagents, with short threads and a handoff at the end.

A representative test gave me a reason to stop. The reported result missed four of ten critical facts from one historical orchestrator session. It also retained stale early statements while omitting later outcomes. A statement that nothing had been pushed could survive after the work had been pushed.

That tested version never became the live configuration. The misses came from the experiment, not from evidence that my normal sessions had lost those facts. It was still enough to reject the proposed change for my workflow.

Keep the useful state intact

The test reduced the context substantially, but the reduction percentage didn’t answer the question I cared about. An agent resuming work needs to know which decisions are settled and what has already happened. Losing an outcome can send it back through work that is complete or make it act on an obsolete assumption.

I chose to stop iterating on compaction and disabled the plugin. I could continue working on shorter stages and explicit handoffs without adding another system to recover the facts those sessions needed.

I kept the narrower classification work moving. Asking a model a bounded question and handling an uncertain answer in code was a separate use case; the compaction result didn’t settle whether that was useful.

The same day’s workflow review also pushed the measurement work forward. Task usage needed to come from the exact session involved, and recording the same completion twice needed to leave one outcome. The implementation report said those changes were in place, with focused checks and an independent review.

The effect on real work was still to be assessed. I wanted to look at subsequent tasks before claiming an efficiency gain.

Written by . Follow via RSS ↗