Agents—JUN 14, 2026 · 3 min read
The empty context window
Context engineering is usually framed as the art of what to put in. The harder discipline — and the one that separates working agents from flailing ones — is what to leave out.
There is a moment, before the first token arrives, when a language model is capable of anything it will ever be capable of. The weights sit there holding every competence they hold — every language, every API, every style of argument — in superposition. Then the context starts to fill, and the model begins the long process of becoming less.
That is not a criticism. Becoming less is the job. A prompt is an act of collapse: from everything the model could say, toward the narrow thing you need it to say. But once you see conditioning as subtraction, the craft of context engineering inverts. The question stops being “what does the model need to know?” and becomes “what am I making it impossible for the model to do?”
Every token is a commitment
Attention is a budget in the strict sense: softmax normalizes to one. Whatever probability mass your debugging transcript from four turns ago is absorbing is mass not spent on the actual question. Long before you hit the context limit, you hit the relevance limit — the point where the marginal token is displacing attention rather than informing it.
Anyone who has operated agents in production has watched the failure mode this produces. An agent thrashes for fifty turns, filling its window with error logs, retries, half-remembered plans. And the fix — embarrassingly often — is to throw the history away. Fresh context, one clean paragraph describing the actual state of the world, and the same model that was flailing solves the task in three moves.
We keep rediscovering this and keep being surprised by it, because it offends our intuition that memory is an asset. For agents, memory is a liability that occasionally pays dividends. The transcript is not the state. The transcript is a record of how you got to the state, and records belong in archives, not in working memory.
Compaction is a spiritual practice
The practical toolkit is well known by now: summarize old turns, externalize durable facts to files or stores, retrieve on demand, keep the window for the live task. Every serious agent harness converges on some version of it.
What’s less discussed is that compaction is a lossy judgment, not a compression algorithm. When you summarize forty turns into six sentences, you are deciding what the past was about. Do it badly and the agent inherits your confusion at higher density. Do it well and something almost paradoxical happens: the agent gets smarter as it forgets, because what remains is only the load-bearing truth.
A useful discipline for writing summaries, human or machine: every sentence must change a future decision. “We tried the streaming API and it 404s on batch endpoints” survives. “We had a long discussion about approaches” does not. If no downstream choice depends on a sentence, the sentence is nostalgia.
Designing for emptiness
The deepest version of this principle shapes the system before any prompt is written. The teams whose agents work are the ones who design environments in which little context is needed: tools with self-describing errors, state that can be re-derived rather than remembered, tasks decomposed so each subagent needs one paragraph of briefing rather than the whole saga. A subagent spawned with a clean window and a precise brief is not a workaround for small context limits. It is the correct architecture even with infinite ones, for the same reason you don’t onboard a new engineer by reading them the company’s entire Slack history.
There is a line attributed to the shakuhachi tradition: the music is not in the flute. The bamboo contributes nothing but a shaped emptiness for breath to move through. It would be too neat to say the intelligence is not in the context — plainly some of it is. But the usable intelligence, the part that shows up when a task is on the line, lives in the shape of what you left out. An overfull window is a flute packed with everything the carpenter knew about wood.
Hold the transcript loosely. Keep the window clean. The model already contains multitudes; your job is to stop most of them from talking at once.