Stage 08 of 11 · Multi-Agent Systems · 3 min · Reviewed Aug 2026
Multi-Agent Systems: Supervisors, Subagents, Parallel, A2A
Specialists help when they produce different artifacts. A supervisor routes work and stops the run. Subagents, parallel calls, workflows-as-tools, and A2A (agent-to-agent protocols) are how you coordinate without a crowd of nicknames.
- Supervisors
- Subagents
- Parallel calls
- Workflows as tools
- A2A
What you learn
How to coordinate multiple specialized agents with contracts, not theater.
Roles need contracts
A search agent should return hits. A writer should return a draft. If both only chat, you have one agent with nicknames. Typed artifacts on the handoff are how orchestrators earn their keep.
Write the schema before you write the second persona. If you cannot name the object that crosses the boundary, do not split.
{
"artifact": "search_hits",
"query": "agent tool timeout",
"items": [
{ "url": "https://example.com/trace", "claim": "Timeouts belong on every network tool." }
]
}Visual
Typed Artifact Handoffs
Each agent produces a typed artifact that the next agent or the supervisor can validate. Without a schema on the boundary, you have nicknames, not roles.
Supervisor and subagents
A supervisor (orchestrator) holds the goal, assigns tasks, validates results, and decides when to stop. Subagents do not see the whole world — they see a scoped task and limited tools. That is the same idea as least privilege.
Dynamic agents (spawn a specialist at runtime) are powerful and easy to abuse. Cap the spawn tree. Log parent/child. Kill orphans.
Visual
Supervisor–Subagent Topology
The supervisor holds the goal and routes tasks. Subagents have scoped tools and limited context — least privilege applied to AI.