Yesterday I asked GitHub Copilot a 12-character question: “Who are you?”
By the time the request reached my local Qwen 3.8 Flash Next model on a DGX Spark, it had grown to 107,527 serialized characters—26,504 input tokens. So I opened the hood.
Inspecting the calls
You can inspect the same machinery in VS Code:
- Open the Command Palette (Cmd+Shift+P).
- Run Developer: Show Chat Debug View.
You’ll see the model calls Copilot makes around a chat request—not only the call that produces the final answer.
First, there are helper calls: modelList, title generation, and progress-message generation. modelList appears to retrieve the models available to Copilot. The title call turns “Who are you?” into “Chatbot introduction.”
Then come the progress-message calls—three in this trace, even for this tiny prompt. One asks the model for phrases such as “Polishing your code,” “Tuning the syntax,” “Refining the logic,” and “Sprucing up the script.” Harmless, but still extra inference.
The mammoth request
The mammoth request is copilotLanguageModelWrapper. Its prompt-bearing payload was 107,527 serialized characters. The endpoint reported 26,504 input tokens. The visible question itself—“who are you?”—was just 12 characters.
Two sections dominate the payload:
- 71 skill descriptors: 40,833 chars
- 25 callable tool schemas: 40,005 chars
Together, that’s about 75% of the serialized input—before counting agent instructions, memory, editor rules, or the actual question.
Also bundled in:
- User memory: 11,747 chars
- 14 agent descriptors: 4,193
- Editing instructions: 2,307
- Core agent instructions: 2,258
- Memory policy: 1,931
- Tool-use instructions: 1,583
- Edit reminders: 588
Question: 12.
Explore the interactive minimap
I turned the request into an interactive minimap. Click any section to inspect a public-safe example of what it contains:
https://copilot-prompt-anatomy.juanan970788.chatgpt.site
This context is assembled by the agent harness; it isn’t specific to Qwen.
The takeaway
The takeaway isn’t “agents are bad.” Tools, skills, and memory make them useful. But context has a cost. Leave everything enabled, and a small prompt can carry tens of thousands of unrelated tokens.
Treat agent context like dependencies: audit it, trim it, keep it intentional.