CCA-F: Claude Certified Architect — Foundations

By Anthropic · Question Mode

QuestionQ71

Context Management & Reliability

You are building developer productivity tools with the Claude Agent SDK. The agent helps engineers explore unfamiliar codebases, understand legacy systems, generate boilerplate code, and automate repetitive tasks. It relies on the built-in tools (Read, Write, Bash, Grep, Glob) and also integrates with Model Context Protocol (MCP) servers.

An engineer asks the agent to identify untested code paths in a legacy payment processing module that spans 45 files. After the agent has read the first 8 source files, its responses start becoming noticeably less accurate: it is forgetting code patterns discussed earlier, and it still hasn't located all the relevant test files or traced the critical payment flows.

What is the most effective approach to complete this investigation?

Explanation

Delegating focused sub-questions to subagents lets each one explore with its own clean context window (e.g., one subagent locates all test files for payment processing, another traces refund flow dependencies), while the main/orchestrating agent retains only the synthesized findings and high-level understanding rather than the raw content of every file. This prevents the context degradation caused by accumulating large amounts of file content in a single context window and allows the remaining discovery work to proceed effectively and in parallel, which directly addresses both symptoms described: declining accuracy from context overload and incomplete discovery of test files and payment flow traces.

Community Discussion

No comments yet. Be the first to start the discussion!