A practical framework for deciding when to implement AI agent functionality as a tool versus a subagent. Tools execute deterministic code (API calls, database queries, calculations) and share the orchestrator's context window, while subagents run their own LLM reasoning loops in isolated contexts. The article presents a three-question decision framework: Is the task execution or reasoning? Does intermediate work matter to the orchestrator? Can the task run independently? It warns against over-engineering by defaulting to subagents prematurely, noting they add latency, cost, and debugging complexity. The recommended approach is to start with a single agent and well-designed tools, only introducing subagents when they provide clear architectural advantages like parallel execution, context isolation, or specialized tool sets.
