When agents need to work together, match the coordination topology to the task: chain for sequential phases, star for parallel decomposition, mesh for exploration. When agents need to challenge each other, design for productive tension — assign conflicting mandates and require explicit engagement with opposing positions. The minority report from an adversarial agent often contains the most valuable content; preserve it rather than averaging it away.
A single agent reasoning alone hits an accuracy ceiling; multiple agents reasoning together hit coordination ceilings. Adding agents improves parallel tasks but degrades sequential ones, and intensive interaction erodes the distinct perspectives that made coordination worthwhile. The designer must choose between collaboration topologies and adversarial structures without clear guidance on when each applies.
Multiple agents with distinct personas need to work together — either collaboratively on a shared task where each agent contributes a phase or domain, or adversarially where structured disagreement surfaces better answers. The designer must choose a coordination topology, assign agent identities, and specify how agents maintain distinct personas under intensive interaction.
For collaborative coordination, choose topology to match task structure:
Assign each agent an explicit purpose, constrained tool set, and permission scope. The supervisor-subordinate variant features a supervisor that holds the whole-task view while specialists maintain depth within their domain.
For adversarial deliberation, design for productive tension: assign agents conflicting mandates, different information, or opposed analytical frameworks. Require explicit agreement/disagreement with justification — agents must respond to each other’s positions, not just state their own. Build a convergence gate that forces crystallization after sufficient deliberation rounds.
Design principles from ICLR 2025 research: maximize reasoning strength with best-available models; use balanced heterogeneous teams with diversity across analytical stances; require non-trivial initial disagreement (moderate disagreement achieves best performance); enforce explicit deliberation with a protocol-specified engagement requirement.
For identity maintenance under coordination, use explicit behavioral rules rather than descriptive persona documents. The SOUL.md approach — a file containing explicit behavioral constraints — resists convergence under interaction pressure better than identity-assertion personas that describe what an agent “is like.” Each agent needs a distinct identity with explicit scope constraints that prevent it from being pulled into another agent’s domain.
Collaborative coordination achieves scale effects on tasks where single-agent performance is limited by context width or reasoning depth — but only if topology matches task structure. Mismatched topology (sequential task in star architecture, parallel task in chain) produces degradation rather than improvement.
Adversarial deliberation produces higher-quality outputs on judgment tasks, but the minority report must be preserved and surfaced rather than averaged away. A convergence mechanism that simply produces the majority view discards the deliberation’s value.
Identity maintenance through explicit rules increases the cost of legitimate persona evolution. When an agent’s role changes, the rule set must be updated explicitly — implicit drift is blocked. This is the intended trade-off.
Gains plateau beyond 4 agents and may reverse. The design question is not “how many agents?” but “what is the minimum number of agents needed to cover the required reasoning stances?”
ICLR 2025 and subsequent multi-agent research identified five structural requirements for effective groups: hierarchy, specialization, division of labor, structured disagreement, and convergence mechanisms. Groups missing any of these degrade toward either unproductive debate (without structure) or groupthink (without disagreement).
Mitsubishi Electric (January 2026) announced multi-agent AI using an argumentation framework to automatically generate adversarial debates among expert AI agents, enabling “rapid expert-level decision-making with transparent reasoning.” The argumentation framework is the convergence mechanism.
Google/MIT research on scaling agent systems: centralized coordination improved parallelizable tasks by 80.9%; sequential reasoning tasks degraded by 39-70% with any multi-agent approach; tool-heavy tasks pay 2-6x efficiency penalty; errors amplify up to 17x without checkpoint mechanisms.
The Structured Disagreement Through Persona Review pattern in this garden instantiates adversarial deliberation with historical thinker personas organized into polarity pairs (Socrates/Feynman, etc.) with a three-round protocol: independent analysis, cross-examination, synthesis. The Groundskeeper-Gardener commission architecture uses the supervisor-subordinate collaborative topology.
extracted_from::[[Persona and Agent Personalities]]↑