Multi-Agent Architectural Patterns¶
The term "multi-agent" is applied to LLM systems with substantially different architectures and properties. Sequential pipelines, role-playing prompts within a single context, orchestrator-worker hierarchies, and concurrent conversational architectures are all described with the same label. A taxonomy based on coordination mechanism — not task type — is needed because the properties of a system (failure modes, coordination requirements, relationship to human authority) differ depending on which pattern is actually implemented.
Ten Patterns¶
The patterns below are composable building blocks, not mutually exclusive architectures. Production systems routinely combine elements from multiple patterns. They are classified along four dimensions: concurrency (whether agents operate simultaneously), context isolation (whether agents access each other's internal state), human participation mode, and turn-taking mechanism.
Pattern 0: Single-Agent Role Simulation (Baseline)¶
A single LLM generates sequential responses attributed to multiple named roles within a unified context. Not a multi-agent architecture — all "agents" share one context window, one set of model weights, one inference process. Shanahan et al. (2023) develop the implication: the model samples from a "superposition of simulacra" rather than switching between committed character states. Included as the baseline against which true multi-agent patterns are measured.
Pattern 1: Autonomous Loop¶
One agent, one task queue, self-directed iteration. Known failure modes: unbounded task generation, context drift, goal misalignment accumulation.
Pattern 2: Sequential Pipeline¶
Tasks flow through agents in a fixed order. Appropriate when task decomposition is well-understood and stable.
Pattern 3: Orchestrator-Worker¶
A central coordinator decomposes tasks and delegates to specialists. The most common production pattern for complex task automation (Wu et al., 2024).
Pattern 4: Role-Based Crew¶
Agents have fixed roles, goals, and personas; a predefined process governs interaction. Works when workflows map to organisational role structures (Hong et al., 2024).
Pattern 5: Conversational Group Chat¶
Multiple agents share a message history and take turns. Speaker selection determines order. Supports emergent reasoning but requires careful turn management.
Pattern 6: Handoff/Transfer¶
One agent explicitly transfers control to another; only one is active at a time. Serial by construction.
Pattern 7: Personal Assistant with Heartbeat¶
A single agent with persistent identity ("soul") and scheduled autonomous invocation ("heartbeat"). The heartbeat enables proactive monitoring without waiting for human input — the agent can act without being asked.
Pattern 8: Organisational Hierarchy¶
Agents arranged in an org chart; top-down task assignment with goal cascading. Agents receive assignments and return results without conversational interaction.
Pattern 9: Shared Room / Concurrent Multi-Agent¶
Multiple agents with separate identities share a communication space and operate concurrently. Each agent maintains an isolated context window and sees only what has been communicated into the shared space. Human participants are first-class members. This is the only pattern combining concurrency, per-agent context isolation, and human peer participation.
Taxonomy Summary¶
| Pattern | Concurrency | Human Role | Context Isolation |
|---|---|---|---|
| 0: Role Simulation | No | Initiator | None |
| 1: Autonomous Loop | No | Observer | N/A |
| 2: Sequential Pipeline | No | Initiator | Partial |
| 3: Orchestrator-Worker | Partial | Approver | Partial |
| 4: Role-Based Crew | Partial | Initiator | Partial |
| 5: Group Chat | No | Peer | Shared |
| 6: Handoff/Transfer | No | Configurable | Shared |
| 7: Personal Assistant | No | Peer | Per-agent |
| 8: Org Hierarchy | Yes (tasks) | Approver | Per-agent |
| 9: Shared Room | Yes | Peer/Authority | Per-agent (enforced) |
Why the Distinction Matters¶
The distinction between Pattern 0 and Pattern 9 is the critical one for safety-critical applications. Pattern 0 (single-agent simulation) shares a single context and model — all "agents" have correlated errors. Pattern 9 (concurrent operation with context isolation) provides the structural basis for epistemic-independence.
A single model holds all simulated perspectives in unified attention (see llm-architecture). Information "known" to one perspective is visible to all within the same context. Prompting cannot override this — it is a mathematical property of self-attention. Systems claiming "independent multi-agent analysis" must specify whether they implement Pattern 0 (simulation) or Pattern 9 (structural separation).
Five Red Flags in Proposals¶
- "Independent multi-agent analysis" where all agents use the same base model
- "AI verification" without specifying what enforces separation between original and verification
- "Self-correcting AI" without external feedback source (see self-correction-limitations)
- "Human-in-the-loop" without architectural enforcement mechanism (see governance-gates)
- Pattern 0 described as Pattern 9
Relevance to Safety-Critical Systems¶
-
Pattern selection determines safety properties. The choice of pattern is not an implementation detail — it determines what independence, reliability, and human authority properties the system can provide.
-
Cross-domain applications. In aviation, a Pattern 7 heartbeat monitor can provide ambient flight condition surveillance. In medical settings, a Pattern 3 orchestrator-worker can coordinate specialist diagnostic agents. In oil and gas, a Pattern 9 shared room can support concurrent well monitoring with independent analysis. The pattern should match the operational requirements, not the other way around.
-
Composability. Real systems combine patterns: Pattern 7 agents within a Pattern 9 room, Pattern 3 orchestration within Pattern 4 role structures. Understanding the individual patterns is prerequisite to understanding compositions.