How Code Agents Locate Code: Tracing Claude's Search Loop

Background: Code Location in Code Agents

Code location is a key capability of code agents. Due to LLM context window limits and reasoning degradation with large codebases, agents must accurately find relevant files and functions for a given natural language intent.

Industry Approaches

Two main approaches exist: embedding-based semantic search and agentic tool-driven loops. Modern code agents predominantly use the latter. The article references two external analyses: https://modem.dev/blog/how-coding-agents-read-your-code and https://offnote.substack.com/p/how-coding-agents-find-their-way.

Experiment: Tracing Claude Code Agent Interactions

The author conducted an experiment using Claude with DeepSeek to observe the interaction flow. The test environment: Claude + DeepSeek. The user question: "Which modules use watchdog?"

Interaction Flow

First request: System prompt defines agent types including an "explore" agent for exploration. Tool descriptions include the user-installed fff search tool with its rules and termination conditions.

Second request: The agent generates a session title summarizing the coding session goal.

Third request: Previous context and tool search results are appended. The agent continues reasoning.

Fourth request: The agent executes a grep command; the command output is added to context and sent to the LLM, which then generates the next command via chain-of-thought.

Fifth request: Further search for callers of the identified code.

Sixth request: Command results are fed back to the LLM for continued reasoning.

Key Findings

Search tool configuration matters: The system context included the fff tool; the author notes that the native grep tool might yield better results and plans to test it.

LLM reasoning chain quality is critical: Higher-quality chain-of-thought leads to more accurate search commands.

References

https://zhuanlan.zhihu.com/p/2023131784934687899