Research, plan, implement keeps coding agents out of the dumb zone
A coding agent gets worse the longer a session runs. Research, plan and implement in separate, compacted steps, use sub-agents to control context, and put human effort where mistakes cost the most.
Written by
-
Benedikt Stemmildt Founder & Co-CEO
Published: 5 February 2026 Updated: 2 min read
Your coding agent gets dumber the longer you use it.
That is not a bug. It is how context windows work, and most developers are fighting it the wrong way.
The dumb zone starts around 40%
Around 40% context window usage, quality starts to degrade. If you have MCP servers dumping JSON everywhere, you are doing all your work in the dumb zone.
The naive way to work with an agent is ask, correct, ask, correct, until you run out of context or give up.
The fix: research, plan, implement
- Research. Understand how the system works. Find the right files. Stay objective. Compress what you found into a summary.
- Plan. Outline the exact steps, with file names and code snippets. Make it so clear that the dumbest model could not get it wrong.
- Implement. Execute the plan. Keep context low. Start fresh windows constantly.
The key is frequent, intentional compaction. Compress, start fresh, and stay in the smart zone.
Sub-agents control context, they do not play roles
Dex Horthy puts it in one line: “Sub-agents are not for anthropomorphizing roles. They are for controlling context.”
So stop creating a “frontend sub-agent”, a “backend sub-agent” and a “QA sub-agent”. That misses the point. Fork a context window to do the searching and reading, return a succinct summary, and the parent agent stays lean.
Human effort belongs in research and planning
Horthy’s second point: AI does not replace thinking. It amplifies the thinking you did, and it amplifies the thinking you skipped just as much.
A bad line of code is a bad line of code. A bad part of a plan can be a hundred bad lines of code. A bad line of research, and the whole thing is hosed.
Move your human effort to the highest-impact parts of the pipeline. That is research and planning, not implementation.
Review the plans
Code review exists for mental alignment: keeping everyone on the same page about how the codebase is changing.
With AI shipping far more code, you cannot read it all. But you can read the plans, and that is enough to catch problems early.
The talk behind this: Dex Horthy, “No Vibes Allowed: Solving Hard Problems in Complex Codebases” (opens in a new tab).
Are you fighting your context window, or engineering around it?
Further reading
Written by
-
Benedikt Stemmildt · Founder & Co-CEO
More on Context engineering
- What context engineering looks like in practice The gap between mediocre and strong agent output comes from how you instruct the agent, not from the model. Concrete practices across three layers: prompting, context and verification.
- A good CLAUDE.md describes how to think, not the project The default CLAUDE.md lists the stack and the commands, and it shapes nothing. The useful ones set philosophy, rules, protocols and principles for how the agent works.
- Most agent failures are context problems, not prompt problems Prompt engineering is how you ask. Context engineering is what the model sees. Eight practices, the symptoms of a badly built context and a four-question audit for the window.
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 QAWant to check one team first? The self-check: twelve yes or no questions, three minutes, instant result, no account. Take the self-check