CCAR-F scenarios › Developer Productivity with Claude · Question 36 of 60
Scenario 4 · Developer Productivity with Claude
Your team is building internal developer tools on the Claude Agent SDK. The agent helps engineers find their way around unfamiliar or legacy codebases, generates boilerplate, and automates repetitive chores. It works through the built-in Read, Write, Bash, Grep, and Glob tools, plus connections to Model Context Protocol (MCP) servers.
Claude Code Configuration & Workflows
An engineer asks the agent to migrate a legacy payments service off a deprecated HTTP client used in about 50 files, where retry and timeout handling differ between modules. In past attempts, the discovery phase filled the conversation with file contents, and by the time the agent designed the approach it was contradicting findings from early in the session. How should this task be run?
Show answer and explanation
Answer: D — Use plan mode, delegating verbose discovery to the Explore subagent so only summaries return, then switch to direct execution to apply the planned migration.
A migration across about 50 files with behaviour that varies by module is a plan-mode task. The Explore subagent keeps verbose discovery output out of the main context and returns summaries, which prevents the degradation seen before. Planning first and then executing directly is the combination section 6 describes. (a) risks expensive rework when the variations show up late. (b) compacts only after the damage is done. (c) assumes the right design is known without exploring the code.
More from this scenario
- A new engineer reports that the agent never adds the team's required license header or…
- An engineer asks the agent to "add comprehensive tests to the legacy inventory service,"…
- An engineer spent a long session having the agent analyze a legacy authentication…
- Your Agent SDK loop for the "automate repetitive tasks" feature sometimes stops halfway…
- An engineer asks the agent, "Where is calculateLateFee called in this legacy billing…