What is a sub-agent?
A sub-agent (or specialist agent) is a domain-specific worker inside a multi-agent AI flow, operating with its own model provider, instructions, tools, and skills. Unlike monolithic agents, sub-agents process dedicated workflow steps independently under a primary coordinator agent.
- Independent Tool & Skill Execution: Sub-agents load only the relevant tools needed for their specific task, reducing prompt overhead and enforcing security boundaries.
- Atomic Handoff Routing: Primary orchestrator agents delegate or transfer user requests to sub-agents via structured routing graphs.
- Isolated Execution State: Each sub-agent maintains its own step budget and execution trace for auditability and state replay.
Sub-agents are where actual workflow execution happens. In Anter's architecture, the root agent handles triage while every domain response is generated by a sub-agent.
Instruction precedence
The root's shared instructions are merged into whichever sub-agent is currently active, so they always exert influence. But when root and sub-agent instructions conflict, the sub-agent's own instructions take priority. Write them assuming they win.
The practical rule: put agent-wide policy in root instructions; put anything behaviour-critical in the sub-agent when the two could disagree.
One job per sub-agent
The most reliable configurations give each sub-agent a narrow, well-defined responsibility. Overlapping responsibilities are the usual root cause of ambiguous routing — when two sub-agents could both plausibly own a request, the model's routing intent becomes unstable and the runtime starts issuing correction nudges.
Reachability
A sub-agent that no other sub-agent can hand off to will never run, even though the design validates cleanly. Every sub-agent should be reachable from the default one. This is one of the few multi-agent bugs that is invisible until you inspect the routing graph as a graph rather than as a list.
Cross-agent sub-agents
A sub-agent can be a reference to a different, externally-linked agent. These behave differently in one critical way: they can only ever be delegation targets, never transfer targets.