The Context Window and Its Limits
The context window is the limit on how much information a model can process at once. It is measured in tokens rather than characters, and a single character can consume more than one token in some languages.
What Counts Toward It
The window covers output as well as input.
| Included | Example |
|---|---|
| System instructions | Role setup, rules to follow |
| Conversation history | Everything exchanged so far |
| Tool definitions | The description text for available tools |
| Tool results | Search output, file contents |
| Generated output | The response about to be written |
The commonly overlooked entries are tool definitions and tool results. Twenty tools consume a meaningful share through their descriptions alone, and a tool that reads whole files consumes heavily in a single call.
The Limit Is Not a Budget to Spend
Even on a model with a large window, filling it is a poor strategy. A model treats everything in context as evidence, so more unrelated material lowers the relative weight of what matters.
Little information passed
→ the instruction and source material land clearly
Information near the limit
→ technically processable, but the important instruction is relatively buriedThe technical limit and the range where quality holds are different numbers.
Behavior Near the Limit
Past the limit, most implementations either truncate older history or return an error. When truncation happens, the important instruction given first can be the first thing removed.
Three responses apply.
- Summarize — compress older history down to decisions and constraints
- Drop — remove consumed intermediate results and full logs
- Split — divide the work and delegate to a subagent with its own window
Summary
- The context window limits how much can be processed at once
- It covers tool definitions, tool results, and generated output, not just input
- Filling it lowers the relative weight of what matters and destabilizes judgment
- The responses are summarizing, dropping, and splitting the work