Skip to content

Blog

Million-token windows did not remove the dumb zone

Agent quality degrades without warning above roughly 40 percent context fill, the zone Dex Horthy named. Windows grew to a million tokens and the failure mode came along, so starting bare metal is still the habit that keeps a session out of it.

Written by

Context engineering

Published: 27 August 2026 Updated: 2 min read

An agent does not tell you when its context is too full to think straight. It keeps answering in the same confident voice, and the answers get worse.

The dumb zone starts around 40 percent

Dex Horthy has a name for where that happens. He calls it the dumb zone, and he puts the line at roughly 40 percent utilization. Below it an agent stays sharp. Above it the quality drops while the agent keeps talking as if nothing had changed.

That is more dangerous than hitting the hard limit. The limit throws an error you can see. The dumb zone throws no error. The agent gets distracted by context that has no bearing on the task, and you only find out by reading the output.

Bigger windows kept the failure mode

Here is the part that should have dated by now and has not. When he said it, the obvious answer was bigger windows. Windows got bigger, and million-token windows are ordinary now. The failure mode came along.

One lab’s own model card, for its own million-token model, reports 26 percent on its long-context recall benchmark at the full window. Another states in its engineering documentation that recall accuracy falls as the window fills. That is the vendors themselves saying the window is not the capacity.

So the labs did go after the degradation: compaction, context editing, memory that lives outside the window. They shipped all of it in the same releases as the bigger windows, not instead of them. Do not wait for a window size that ends this.

Tool definitions fill the window first

The usual culprit on your side is tool definitions. Every MCP server you connect loads its whole catalog into context before the first prompt. Connect three or four and the window is already crowded with definitions the task will never touch. The agent reads all of it.

Start bare metal

The fix is a habit, not a setting. Start bare metal. Open a session with as little loaded as possible, then add only what the task in front of you needs. Keep MCP servers for the few cases where a CLI cannot do the job, and reach for the CLI everywhere else. Watch the fill percentage the way you watch a build status.

Codebase hygiene gets a lot of attention. Context hygiene is the same craft, one layer up.

When did you last look at how full your agent’s context was before it started work?

(Sources: Dexter Horthy, “No Vibes Allowed: Solving Hard Problems in Complex Codebases” (opens in a new tab). Model card figure: Gemini 3.1 Pro model card (opens in a new tab).)

Further reading

  1. Who owns the file every agent reads first Since Claude Code 2.1.277 reads AGENTS.md where no CLAUDE.md exists, one instruction file can serve every coding agent on a team. The file was never the hard part: the instructions are a team decision and need one owner, so merge the files into one AGENTS.md and name who owns it. 1 min readShort post
  2. Feed the factory the decisions nobody wrote down In most organizations the knowledge that matters most is decided out loud in meetings and never written. Five concrete steps route it into the context layer the factory reads. 2 min readShort post
  3. AI tool fatigue is a missing context layer Teams count their AI tools when the list that decides is the second one: skills, commands, CLAUDE.md entries and context files. Where that list is rich, the tool list shrinks on its own. 1 min readShort post

All 16 posts on Context engineering

ClientsCase study: A large DIY retailer in Germany

Scaling needs named owners

Coding agents scale across teams when named people own the platform, the team practice, the guardrails, and the measurement. Rollout & scale is where that ownership is built and handed to internal owners, in your stack.

Training on this topic: Agentic Engineering for Developers and QA
See Rollout & scale

Want to check one team first? The self-check: twelve yes or no questions, three minutes, instant result, no account. Take the self-check