React Basics for Backend Engineers

Scenario: Developer Productivity with Claude You are building developer productivity tools using the Claude Agent SDK. The agent helps engineers explore unfamiliar codebases, understand legacy systems, generate boilerplate code, and automate repetitive tasks. It uses built-in tools (Read, Write, Bash, Grep, Glob) and integrates with MCP servers. Your developer productivity agent uses a coordinator that delegates codebase exploration tasks to subagents. Engineers invoke the agent by describing a goal in natural language, such as "explain how the authentication middleware works" or "find all places that call the payment service." In testing, you observe that subagents frequently produce shallow or off-target results: they explore tangentially related modules, miss the core logic engineers care about, and occasionally ask clarifying questions mid-task that the coordinator cannot answer. Reviewing the coordinator's prompts to subagents, you notice they look like this: "Explore the authentication system and report back what you find." What is the most likely cause of these subagent failures?