What is Multi-Agent?
Multi-Agent Systems are AI application architectures in which two or more bounded agents coordinate through explicit task, message, state, authority, and completion contracts.
Quick Facts
| Full Name | Multi-Agent Systems |
|---|---|
| Created | Concept from 1980s AI, LLM-based from 2023 |
How It Works
Choose agents from task structure
Use multiple agents when independent branches can run concurrently, one capability needs a distinct permission boundary, a specialist needs isolated context, or organizational ownership requires a stable service contract. A long task, many steps, or several personas does not by itself justify multiple agents. Establish a single-agent or deterministic-workflow baseline and compare accepted outcomes, latency, cost, and operator burden before expanding the topology.
Topology and control transfer
Manager-as-tools keeps one coordinator responsible for the user-facing result while specialists return bounded artifacts. A handoff transfers active control and usually changes the instructions and context governing the next turn. A deterministic workflow fixes sequence, branches, or fan-out in code; peer or mesh designs distribute control further. These are distinct decisions. Each transition should identify the owner, destination, allowed reason, context subset, remaining budget, and return or termination condition.
State, artifacts, and side effects
Assign one authoritative writer for each mutable state field and exchange immutable, schema-versioned artifacts rather than shared prose. Every task and effect needs a run ID, task ID, principal, deadline, attempt, and idempotency key. Treat peer output as untrusted and reauthorize every tool call at the execution boundary. A durable checkpoint records logical progress, but external writes still require reconciliation with the system of record and an explicit unknown-outcome path.
Failure handling and evaluation
Propagate cancellation and deadlines, cap delegation depth and total work, reserve budget for synthesis, and define how partial results are accepted or discarded. Detect no-progress loops, duplicate work, contradictory artifacts, unavailable workers, and late results. Evaluate final state and user value first, then inspect routing accuracy, coverage, handoff loss, repeated effects, recovery, safety, token use, and p95 latency. Agent-level scores alone cannot prove the coordinated system works.
Key Characteristics
- Introduces multiple bounded execution units with explicit ownership
- Separates coordination topology from sequential, parallel, routing, and loop control flow
- Uses versioned task, artifact, handoff, and terminal-state contracts
- Requires least-privilege authorization at every agent and tool boundary
- Needs global budgets, cancellation, idempotency, reconciliation, and loop detection
- Must be evaluated against a simpler baseline on complete system outcomes
Common Use Cases
- Running independent research branches in parallel before evidence-aware synthesis
- Separating implementation, testing, and approval identities in software delivery
- Delegating domain tasks to separately governed services through stable contracts
- Using an independent reviewer when correlated self-review is an identified risk
- Coordinating long-running work that needs checkpoints, bounded retries, and human escalation
Example
Loading code...Frequently Asked Questions
What is the difference between multi-agent systems and a single AI agent?
A single agent owns one main context and control loop. A multi-agent system divides work among bounded agents and adds contracts for delegation, context transfer, state, authority, and completion. The extra architecture is justified only when isolation, parallelism, or ownership improves measured outcomes enough to offset more calls and failure paths.
How are manager-as-tools and handoff patterns different?
With agents as tools, a manager retains task ownership, sends bounded subtasks, and synthesizes the final response. In a handoff, the selected specialist becomes the active owner for the next part of the interaction. This changes which prompt, context, guardrails, and return path apply, so the transfer must be explicit and auditable.
Should agents share one memory or context?
Not by default. Pass the minimum task-specific context and immutable artifact references needed by the receiving agent. Shared mutable state creates races, hidden coupling, privacy exposure, and unclear ownership. Keep business facts in a source of truth, assign one writer per field, and record versioned changes through checkpoints or events.
What types of tasks are best suited for multi-agent systems?
Good candidates have independently executable branches, distinct permission or context boundaries, separate domain ownership, or a need for independent evidence. Fixed, predictable steps usually belong in ordinary code or a deterministic workflow. Highly dependent tasks with one shared context may become slower and less reliable when split across agents.
How do you ensure reliability and handle failures in multi-agent systems?
Define task and handoff schemas, per-agent permissions, deadlines, global work budgets, stop conditions, state ownership, and idempotency for effects. Retry only operations known to be safe, reconcile uncertain writes against authoritative systems, and escalate conflicts or consequential actions. Evaluate complete outcomes and trajectories, not just final prose.