Skip to content

l2-agent-loop.en.md

Latest commit

 

History

History
45 lines (29 loc) · 1.07 KB

l2-agent-loop.en.md

File metadata and controls

45 lines (29 loc) · 1.07 KB

Layer 2: The Agent Core Loop

中文 | English

Core Question

How does Claude Code keep cycling between model output and tool execution until the task is done?

Run First

python examples/l2_agent_loop.py

This requires DEEPSEEK_API_KEY.

Read First

  • query.ts
  • QueryEngine.ts
  • tasks/

Search Anchors

  • export async function* query
  • async function* queryLoop
  • while (true)
  • tool_result

What To Notice

  • who owns message history
  • why tool results must be appended back into history
  • what actually ends the loop
  • how streamed events relate to the final answer

Demo vs Real Source

The demo shows the minimum loop. The real source also handles streaming events, recovery, cancellation, permissions, task state, and UI event emission.

Questions To Answer

  1. Why is the agent loop closer to a state machine plus event stream than a plain while loop?
  2. If the model returns multiple tool calls, how does Claude Code keep the context coherent?
  3. What belongs in QueryEngine and what belongs in query.ts?