02 · Primer
In one paragraph
A multi-agent system is a set of specialised AI agents that coordinate to complete a shared goal faster and more reliably than a single monolithic agent. Split responsibilities, use shared state contracts, and keep one orchestration layer for traceability and conflict resolution.
On this page4 sections
A multi-agent system is a set of specialised AI agents that coordinate to complete a shared goal faster and more reliably than a single monolithic agent.
Five-agent architecture diagram
A coordinator shares tasks, each agent returns artifacts, and a policy agent validates before handoff.
- Spark — planner agent that converts goals into executable task graphs.
- Pixel — design and UI agent that composes layouts and visual assets.
- Ledger — data and state agent that tracks dependencies, budgets, and artifacts.
- Sentinel — risk and policy agent enforcing validation, auth, and compliance checks.
- Rook — execution agent handling deployment steps, retries, and handoffs.
Typical handoffs:
- Spark → Pixel: layout directives
- Pixel → Ledger: asset and state sync
- Ledger → Sentinel: policy check request
- Sentinel → Rook: approved execution
Communication patterns
Three topologies cover most production systems. Pick based on how much coordination overhead the workflow can absorb.
- Hub-and-spoke coordination. A coordinator assigns tasks to specialist agents and merges outputs into a single execution plan.
- Peer-to-peer negotiation. Agents exchange proposals directly when tasks overlap, then commit a conflict-free version.
- Event-driven broadcast. Agents publish state updates to a shared bus so interested agents react without tight coupling.
Worked coordination example
The five named agents above are an explanatory model, not a production log or a claim that seven client sites launched overnight. A bounded site-concept batch could use the model like this:
- Spark turns the accepted brief into independent concept scopes and explicit non-goals.
- Pixel produces page-system proposals while Ledger records assets, content dependencies, and acceptance criteria.
- Sentinel checks schemas, permissions, and unresolved decisions before any external action.
- Rook runs approved local build and verification steps, preserving failed states for human review.
The output of that sequence is a reviewable concept batch. Deployment, production readiness, client acceptance, and business outcomes require separate evidence from the exact sites involved.
Frequently asked
What's the main advantage?
It distributes work across specialised agents so teams can parallelise complex workflows while keeping quality controls in place.
Do they always need a central coordinator?
Not always, but most production systems use at least a lightweight orchestrator to reduce conflicts and improve observability.
How do you prevent agent conflicts?
Shared state contracts, clear ownership boundaries, and arbitration rules when two agents propose incompatible changes.
When should teams avoid multi-agent design?
If one deterministic pipeline solves the problem, adding multiple agents increases overhead without meaningful payoff.
Further reading