Context Engineering

Context engineering is the work of deciding what the model sees on each turn. It is the highest-leverage thing I know of for a long-running agent, and most teams do it by accident: tool results pile up, old file versions sit next to new ones, a subagent compresses its findings into a paragraph that drops the one fact that mattered. The agent is not confused. Its context is.

The essays under this tag treat the context window as a budget to spend rather than a bucket to fill. Compaction keeps the plan from turn three alive at turn forty. Offloading moves a 40,000-token tool result to a store and leaves a handle behind. Pruning removes observations that have gone stale, and a return contract makes the subagent boundary explicit instead of lossy.

Start with context compaction for long-running agents, which is the foundation the rest build on. Then stale tool observations and drift, which shows that this is a correctness problem and not just a token bill, and the subagent return contract for the multi-agent version of the same mistake.