subagents

Overview

Subagents are a workflow pattern in which a primary AI system delegates parts of a larger task to multiple helper agents, often in parallel. Instead of forcing one model instance to hold the entire problem, the main agent breaks work into smaller units, assigns them to specialized or scoped subagents, and then combines their outputs into a final result. This pattern is especially useful for coding, research, migration projects, and other multi-step tasks where a single model can become bottlenecked by time, context window limits, or task complexity.

For AI Product Managers, subagents matter because they offer a practical design pattern for improving speed, context management, and execution reliability in agentic products. They can help products handle long-running work, reduce top-level context overload, and enable more modular agent behavior. As PMs evaluate AI copilots, coding agents, and workflow automation tools, subagents are becoming an important architectural concept for shipping systems that scale beyond single-prompt interactions.

Key Developments

• 2026-01-23: Cursor introduced subagents for parallel task execution, highlighting faster execution, improved context usage, and support for long-running tasks.
• 2026-02-22: Boris Cherny described using subagents with Git worktrees to parallelize large codebase migrations, assigning each agent a subset of folders while a main agent handled merge conflicts.
• 2026-03-17: OpenAI Codex announced general availability of subagents and support for custom agents, extending patterns similar to Claude Code's subagents and TOML-defined agent configurations.
• 2026-03-18: Simon Willison explained subagents as a way to manage LLM context limits by splitting oversized tasks into smaller agentic components, preserving top-level context and improving outcomes.

Relevance to AI PMs

• Designing around context limits: Subagents give PMs a concrete pattern for handling workflows that exceed a model's effective memory. Instead of enlarging prompts indefinitely, teams can decompose jobs into smaller agent tasks and preserve the main agent for orchestration.
• Improving latency and throughput: Parallel subagent execution can reduce end-to-end task time for research, code generation, QA, and migration workflows. PMs can use this pattern to define product experiences where multiple task branches run at once rather than serially.
• Creating more controllable agent systems: Subagents enable role-based task design, such as explorer, worker, or reviewer agents. This helps PMs specify clearer boundaries, permissions, and success criteria for each step of an agentic workflow.

Related

• claude / claude-code: Claude Code helped popularize named subagent patterns such as explorer, worker, and default agents.
• openai-codex: Codex expanded subagent support and paired it with custom agents, making the pattern more mainstream for developer workflows.
• custom-agents: Subagents are often implemented as customizable agent roles with different tools, prompts, or scopes.
• llm-context-limits: One of the strongest motivations for subagents is reducing pressure on a single model's context window.
• toml: TOML is used in some ecosystems to define custom agent behavior and configuration.
• git-worktrees: Git worktrees pair naturally with subagents for parallel code changes across separate branches or folders.
• boris-cherny: His migration example showed a concrete high-value use case for subagents in large engineering projects.
• cursor: Cursor's implementation highlighted the product value of parallel execution and long-running task support.