Skip to content

Multi-Agent Functional Roles

A Pattern 9 architecture requires intentional role assignment. The coordination challenges of multi-agent systems are better addressed by agents configured for specific functions than by general-purpose agents. The MAST failure taxonomy (Cemri et al., 2025) identifies absent role specification as a primary failure mode. DyLAN (Liu et al., 2024) found up to 25% task performance improvement from dynamic role assignment.

Core Roles

Adversarial / Devil's Advocate

Challenges claims, seeks counterevidence, resists premature consensus. Directly implements the "questioning attitude" valued in safety culture. Its effectiveness depends on model separation — an adversarial agent on the same base model as the agents it challenges will tend toward the same conclusions despite its role assignment (Chan et al., 2024). For effective adversarial function, the agent requires either a different base model, substantially different fine-tuning, or restricted information access.

Synthesiser

Aggregates outputs across agents; identifies consensus, conflict, and gaps. Requires privileged read access to all agent outputs and high context budget. The synthesiser produces the integrated assessment that the operator evaluates — it is the bottleneck between multi-agent analysis and human decision-making.

Domain Expert

Provides authoritative specialised knowledge; flags out-of-scope queries. Requires specialised fine-tuning or high-quality retrieval-augmented-generation / knowledge-graphs access. The domain expert's credibility depends on its knowledge base quality.

Monitor / Metacognitive

Observes team process rather than operational content; flags stale context, loops, coordination failures. Requires access to team-level metadata (who responded when, how long since each agent's last output, whether outputs are becoming repetitive). Addresses the multi-agent-coordination-failures that content-focused agents cannot detect.

Coordinator / Planner

Structures work; tracks dependencies and progress; prevents duplication. Operates as a peer, not a hierarchical orchestrator — it coordinates among equals rather than commanding subordinates. This distinction is important: the Coordinator should not have authority to override other agents' assessments.

Projection Agent

Dedicated to Level 3 SA: extrapolates current state, flags emerging conditions. May be coupled to physics-based simulation codes or domain models for forward projection. The distinction from other agents is between Level 1–2 SA (perceiving and comprehending what is happening now) and Level 3 SA (projecting where things are heading).

SA Bridge

Translates between the multi-agent system's internal technical communication and human-readable format. Adapts presentation to the recipient's role and current needs. Different human roles (supervisor, approver, authority holder) need different SA support — the SA Bridge provides differentiated handoffs. See situation-awareness-in-human-ai-teams.

Grounding Agent

Supplies domain knowledge or commonsense facts to prevent error loops. In AutoGen's ALFWorld experiments (Wu et al., 2024), a grounding agent that injected commonsense rules (e.g., "you must find the object before you can examine it") improved success rates by 15% over systems without grounding. The grounding agent addresses the repetitive action loop failure mode in multi-agent-coordination-failures by breaking chains of errors before they propagate.

Safeguard Agent

Checks outputs for safety, correctness, or policy compliance before execution. In AutoGen's OptiGuide coding application (Wu et al., 2024), a safeguard agent that reviewed code before execution boosted unsafe-code detection F1 by 8-35%. This maps directly to the governance-gates concept — the safeguard agent is an automated governance gate that enforces checks before consequential actions proceed.

Task Specifier

Converts a human's high-level idea into a concrete, agent-executable task description. Li et al. (2023) introduced this role in the CAMEL framework: a human provides a preliminary idea (e.g., "develop a trading bot"), and the task specifier generates a detailed specification that the working agents can decompose into steps. This role bridges the gap between human intent and the structured prompts that agents require, removing the need for human prompt engineering expertise. In safety-critical contexts, the task specifier's output should itself be subject to human review via governance-gates, since misspecification of the task propagates through all downstream agent activity.

Memory Keeper

Maintains episodic record; surfaces precedents; prevents re-derivation of previously established conclusions. Requires external memory-architectures with retrieval and persistence. Addresses context overflow by maintaining the team's long-term memory outside individual agent context windows.

The Degradation Curve

Adding agents improves system coverage up to a point, beyond which coordination overhead and human cognitive load cause net performance to decline. A 3-agent architecture may provide better effective coverage than a 6-agent architecture if the operator can maintain SA across 3 but not 6. The location of this inflection point is unknown and likely task-dependent. No published study has characterised this curve. Its empirical determination is a prerequisite for sizing multi-agent deployments. See situation-awareness-in-human-ai-teams for the cognitive load analysis.

Role Assignment via Inception Prompting

Li et al. (2023) demonstrated that roles can be assigned through inception prompting — structured system prompts that define each agent's identity, communication protocol, and behavioural constraints. Key elements include explicit role identity statements, role-flipping prohibitions, structured output formatting requirements, and termination tokens. The prompts are symmetric between cooperating agents, which helps maintain protocol stability over extended conversations.

However, prompt-based role assignment has inherent fragility: the four cooperative failure modes documented by Li et al. (role flipping, instruction repetition, flake replies, infinite loops) all represent breakdowns in prompt-enforced role adherence. In safety-critical applications, role enforcement should be architectural (separate model instances, different system prompts, tool access restrictions) rather than relying solely on prompt instructions that the model may violate. See multi-agent-coordination-failures for the full failure taxonomy.

Role Assignment Principles

  1. Model diversity for adversarial roles. The adversarial agent must be on a different base model to provide genuine independence. See epistemic-independence.

1b. Validate role assignments empirically, not by intuition. Liu et al. (2024) found that human-intuitive role selection (choosing agents whose role descriptions match the task domain) sometimes performs worse than random selection. An agent labelled "Doctor" may not contribute most effectively to clinical knowledge tasks if another agent's reasoning patterns (e.g., "Programmer") provide more useful structure. In safety-critical deployments, role effectiveness should be measured through performance data, not assumed from role descriptions.

  1. Disagreement is signal, not noise. Agents with different information reaching different conclusions signals ambiguity in the evidence. Systems designed to suppress disagreement through forced consensus lose the diversity that justifies the architecture.

  2. Roles are composable. A single agent may serve multiple roles (e.g., domain expert + projection agent for a specialised domain). But combining the adversarial role with any consensus-building role defeats the adversarial function.

  3. The MAST taxonomy warns against absent role specification. Cemri et al. (2025) found that multi-agent systems without explicit role definition produce more coordination failures. Role clarity is a prerequisite for effective multi-agent coordination.

Relevance to Safety-Critical Systems

  1. Role selection determines system properties. Which roles are deployed and how they are configured determines whether the system provides genuine independent analysis, effective human support, and robust coordination.

  2. Cross-domain examples. In aviation, the adversarial role maps to the "monitoring pilot" function that challenges the "flying pilot." In medical settings, the synthesiser role maps to the team lead who integrates specialist opinions. In oil and gas, the projection agent maps to the wellbore modelling function that projects future well conditions.

  3. Not all roles are needed for all applications. A simple ambient monitoring system (Pattern 7) needs no adversarial agent. A safety-critical verification system (Pattern 9) needs adversarial, synthesiser, and governance roles at minimum. Scale the role set to the application's requirements.