Multi-Agent Networks: Composing Specialist Agents That Collaborate
Patterns for routing work across specialist agents without losing context, governance, or observability.
A single agent with every tool eventually becomes a swiss-army knife that is hard to instruct, evaluate, and secure. Multi-agent networks split responsibility across specialists—billing, support, scheduling, research—and coordinate them toward a shared goal.
When Multiple Agents Help
Consider a network when:
- Domains have different Action sets and risk profiles
- Instructions would otherwise become a long, conflicting policy document
- Teams own different backends and want separate Environments and releases
- Some paths need stricter HITL than others
If every request needs the same tools and tone, a single well-scoped agent is simpler.
Common Topologies
Router → specialists
A triage agent classifies intent and hands off to a specialist. The specialist owns the tools and returns a result or a structured handoff payload.
Planner → workers
A planner decomposes a task; workers execute Actions in parallel or sequence; the planner synthesizes the answer.
Peer handoff
Agents pass conversation state when the user topic shifts (e.g., support → billing), with an explicit transfer summary.
Handoff summary:
- User: Acme Corp admin
- Goal: Dispute invoice INV-4401
- Already verified: account ID acct_98
- Next Action candidate: open_dispute (requires approval)
Shared Context Without Shared Chaos
Specialists need enough context to act, not the entire history of every peer.
- Pass a handoff packet: goal, constraints, verified IDs, screen state
- Keep secrets out of free-text summaries; rely on Environment credentials
- Prefer structured fields over prose when another agent will call tools
Screen Observer context from the user's current App view is especially useful at handoff boundaries—specialists should see the same page entities the user is looking at.
Governance Across the Network
Each specialist should inherit appropriate controls:
- Billing agent: HITL on refunds and credits
- Support agent: autonomous lookups, approval for account changes
- Ops agent: Tool Monitoring with tight anomaly thresholds
Do not let a permissive triage agent call sensitive Actions "because it has the conversation." Route to the specialist that owns those Actions.
Observability for Multi-Agent Runs
A user-visible conversation may span multiple agent IDs. Stitch Traces with a shared conversation or correlation ID so you can answer:
- Which specialist handled which step?
- Where did handoffs lose information?
- Which agent's Actions failed?
Without cross-agent Traces, multi-agent systems become harder to debug than monoliths.
Failure and Fallback
Define what happens when a specialist is unavailable or declines a handoff:
- Router retries with a clarifying question
- Fallback specialist with a reduced tool set
- Escalation to a human queue
Document these paths in instructions so agents do not invent silent dead-ends.
Evaluating Networks
Evals should cover:
- Correct routing of intents
- Preservation of critical IDs across handoffs
- No privilege escalation via the wrong specialist
- End-to-end task completion, not only per-agent unit tests
Promote network changes Environment by Environment the same way you promote a single agent.
Conclusion
Multi-agent networks work when specialization is real: different Actions, policies, and owners. Use clear handoff packets, shared correlation in Traces, and HITL on the agents that can cause harm. Composition beats monolith—if you invest in the seams.
