The Context Window Is a Budget

Bigger context windows didn't fix your agent. They hid the failure.

Your agent worked perfectly in the demo. In production it forgets instructions, contradicts itself, and quietly gets worse the longer it runs. So you reach for a bigger model and a bigger window, and nothing improves, because the smooth sentences never sound any worse on the way down. You are not facing a model problem. You are facing a context problem, and almost no one has named it.

Every frontier model degrades as its input grows, measurably and predictably, even far below the token limit. The field calls it context rot, and it is the single biggest cause of agent failures in production. The Context Window Is a Budget treats the fix as an engineering discipline: context engineering, the craft of deciding exactly what your AI agents see, when they see it, and what never enters the window at all. Optimizing what goes into the window beats enlarging it, every time.

Inside, the operating system for building reliable AI agents:

• Context rot, named: why models get worse as input grows, and why you never hear it happen, since the prose stays smooth all the way down.
• The attention budget reframes the window the way a finance team treats money: a scarce allocation you spend deliberately, not a bucket you fill to the brim.
• Write / Select / Compress / Isolate: the four-move taxonomy turned into a working system you can ship this week.
• Why bigger windows hide failures instead of fixing them, and what actually goes wrong inside a long-horizon run.
• Reliability that survives model upgrades: curate what the agent sees, manage its memory outside the window, and stop rebuilding your stack on every release.

Curation, not capacity, is the lever. Read it and you will stop blaming the model and start architecting what it sees, diagnose context rot before your users feel it, and hold your reliability through every upgrade, the one applied AI skill that does not expire in six months. Learn to spend the window.

For applied AI engineers, data scientists, and developers building retrieval augmented generation (RAG) pipelines, agentic RAG systems with rerankers, and long-running agents on the Model Context Protocol (MCP). Part of the Build Agents You Can Trust series, in The Verifier's Library.