What compaction is In the material on coding agents, compaction refers to an agent condensing the conversation and work context it has accumulated — the back-and-forth, the files touched, the plan so far — into a compact summary, so it can keep working within the limits of its context window [^619ac4a8-531a-53a9-9188-912ae7459502]. A long, autonomous task builds up history the model can't hold forever; compaction is the mechanism that lets the agent continue past that point by carrying forward a distilled version of what matters instead of the full transcript. Where it shows up in practice Compaction matters most in the delegation-first workflows described across the coding-agents sessions: you hand an agent a whole task, it plans, edits, and verifies over many steps, and somewhere in that long run its context fills up. When compaction happens, what survives the summarization step becomes the agent's working memory going forward. What to explore next - Notice when an agent's behavior shifts after a long run — compaction is often the reason. - Keep durable state (the plan, key decisions, file list) explicit in the task description so it survives the compression step. - Treat compaction as a reminder that the human's job is planning, orchestrating, and verifying: checking the agent's output matters precisely because its internal context is not permanent.