What is a subagent?
A subagent is a child AI agent that a main agent starts for a side task, with its own context window and instructions, and that returns only a condensed result to the main agent.
A main agent working on a long task accumulates context. Sending a side task, such as researching one question or reviewing one file, to a subagent keeps that work out of the main context. The subagent explores in its own window, and only its conclusion comes back.
Subagents became popular in 2025 for research and coding agents. In 2026 the labs narrowed the advice: they help for specific reasons, and many systems work better with fewer of them.
When a subagent earns its place
Context isolation
A side task that would flood the main context with search results or logs.
Parallelism
Independent pieces of work, such as researching several sources, that can run at the same time.
Conflicting instructions
A step that needs a different role or rule set from the main agent, such as a skeptical reviewer.
What the labs measured
Anthropic's multi-agent research system (June 2025) used a lead agent that spawns parallel subagents. Multi-agent runs used about 15 times the tokens of a chat, so the pattern pays off on high-value, breadth-first work, and Anthropic called coding a poor fit.
Google Research's controlled study (January 2026) found multi-agent coordination helped on parallelisable tasks but cut performance by 39% to 70% on sequential reasoning, and that independent agents without cross-checking amplified errors about 17 times.
The 2026 revision
Anthropic's October 2026 guidance on folding subagents into the main loop found that, for many enterprise agents, one loop with progressively disclosed skills and tools beat orchestrator and subagent designs on quality, cost and latency. It warns against subagents shaped like an org chart and against subagent sprawl, and advises migrating one subagent at a time while comparing evals.
How Motiontic's engineers apply it
Motiontic's engineers start with a single agent loop and add a subagent only for one of the three reasons above. Each subagent gets the minimum context for its job, a fixed format for what it returns, and only the tools it needs, and it is removed if evals show the single loop does as well.
Subagent, agent handoff and tool call compared
| Subagent | Agent handoff | Tool call | |
|---|---|---|---|
| Who keeps control | The main agent | The receiving agent | The calling agent |
| Context | Its own, fresh window | Passed or scoped from the first agent | None, it is a function |
| Returns | A condensed result | Nothing, control moves on | The tool's output |
| Fits | Isolated or parallel side tasks | Moving a conversation to a specialist | A defined action or lookup |
Common questions
What is the difference between a subagent and a multi-agent system?
A subagent is one child agent started by a main agent for a side task. A multi-agent system is any design with several agents, which may use subagents, handoffs between peers, or agents calling each other through a protocol such as A2A.
Do subagents cost more?
Usually. Each subagent runs its own model calls with its own context. Anthropic reported multi-agent research runs using about 15 times the tokens of a chat.
When should subagents be avoided?
For tightly sequential work where each step depends on the last, and when a single agent with well-described skills and tools already performs well. Google Research found coordination cut performance by 39% to 70% on sequential reasoning tasks.
Related
Single Agent vs Multi-Agent
One loop with good tools, compared with several agents splitting the work.
Orchestrator-worker pattern
A lead agent plans, workers run in parallel with their own context, the lead merges.
Agent handoff
One agent passes the conversation to another, which carries on from there.
Progressive disclosure
Show the menu first, load the dish only when it is ordered.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Subagent sprawl
Too many overlapping subagents, each owned by a different team and none by the whole.
Sources
- Anthropic, How the multi-agent research system was built (2025-06-13)
- Google Research, Towards a science of scaling agent systems (2026-01-28)
- Anthropic, When to fold subagents into the main agent loop (2026-10-22)
- Google, Architecting efficient context-aware multi-agent framework for production (2025-12-04)
- Anthropic, Building agents with the Claude Agent SDK (2025-09-29)
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.