Emergent Coordination Protocols Among Heterogeneous LLM Agents

1. Introduction

It has become routine to compose pipelines from agents drawn from multiple vendors --- e.g., a planner from one provider, a code-writer from another, an evaluator from a third. The conventional wisdom is that such pipelines require an explicit communication protocol designed up-front. We study an alternative: what happens if no protocol is specified and agents are only told the high-level task?

The phenomenon is reminiscent of emergent communication studied in multi-agent reinforcement learning [Lazaridou and Baroni 2020], but with an LLM-specific twist: the agents bring strong linguistic priors and substantial in-context reasoning, so convergence is typically faster and the converged protocol is human-legible.

We ask:

• How fast does convergence occur, and to what?
• Do heterogeneous pools converge differently from homogeneous ones?
• Are the converged protocols efficient relative to a hand-designed baseline?

2. Setup

2.1 Tasks

• Scheduling: 5 agents jointly schedule 24 meetings across constraints. Communication is unrestricted text.
• Distributed code review: 4 agents review a 600-line PR sequentially with arbitrary inter-message format.
• Synthetic market: 8 agents trade in a posted-offer market, communicating intent via free-form messages.

2.2 Agents

We sample from 12 model variants spanning four vendors. A pool is a tuple specifying which models occupy which roles. We test 24 pool configurations, 8 of which are homogeneous (all-from-one-vendor) and 16 heterogeneous.

2.3 Metrics

• Convergence time $T_c$: turns until message format stabilizes (Levenshtein distance of consecutive same-role messages drops below 0.15 for 3 consecutive turns).
• Vocabulary entropy $H$: Shannon entropy over message-prefix tokens, computed in a 50-message sliding window.
• Task success: domain-specific (e.g., schedule satisfies all constraints).

3. Results

3.1 Convergence dynamics

Median $T_c$ across all configurations is 7.4 turns; the 90th percentile is 13. Convergence is faster in homogeneous pools ($T_c = 4.8$) than heterogeneous ($T_c = 9.1$), unsurprisingly.

Vocabulary entropy in the converged regime stabilizes at $H = 3.1 \pm 0.4$ bits across pools. For comparison, a hand-designed JSON protocol for the same tasks has $H = 2.7$, suggesting emergent protocols are slightly more verbose but in the same complexity band.

3.2 Convergence regimes

We observe three regimes:

• Vendor-dominant (38% of heterogeneous pools): the protocol of one vendor's model becomes canonical.
• Hybrid (44%): a fused protocol with structural elements from multiple agents.
• Stigmergic (18%): no per-agent protocol; agents communicate via shared scratchpad updates rather than messages.

The stigmergic regime is surprising: it appears most often when one agent in the pool prefers terse outputs and others prefer verbose ones, and tends to be most efficient by token count.

3.3 Task success

No emergent regime is statistically inferior to the hand-designed protocol on more than one metric.

4. Noise-Injection Intervention

In vendor-dominant convergence we worry about protocol monoculture and the loss of redundancy. We test a simple intervention: each turn, with probability $\rho = 0.05$, an agent's outgoing message is rewritten by a small auxiliary model.

def inject_noise(msg, rho=0.05):
    if random.random() < rho:
        return paraphrase_via_aux(msg)
    return msg

Under $\rho = 0.05$, the share of vendor-dominant convergences drops from 38% to 17% with no significant loss of task success ($p = 0.34$). Higher noise levels degrade task success linearly.

5. Discussion

The key takeaway is that under-specified multi-agent systems can self-organize to operating points comparable to hand-designed protocols, at least at the scale tested. This is not an argument against explicit protocols --- which remain superior for safety-critical use --- but it changes the calculus for prototyping and for systems where protocol designers are unavailable.

The stigmergic regime deserves special attention. It transfers state via a shared workspace rather than messages, and it scales better in agent count: at 12 agents, message-based protocols incur quadratic chatter while stigmergic costs grow linearly.

6. Limitations

Our tasks are synthetic; whether the same convergence dynamics hold in production workflows is unstudied. We did not control for prompt engineering effects beyond a fixed system message. Pool size was capped at 8 agents; behavior at 30+ agents may differ qualitatively.

7. Conclusion

Heterogeneous LLM pools coordinate faster and more competently than the literature on emergent communication might suggest. We catalog three convergence regimes and provide a simple lever (noise injection) to reduce monoculture risk. Code, prompts, and traces are released.

References

1. Lazaridou, A. and Baroni, M. (2020). Emergent Multi-Agent Communication.
2. Park, J. S. et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior.
3. Du, Y. et al. (2024). Improving Factuality and Reasoning via Multiagent Debate.
4. Gibney, E. (2025). When AI Agents Talk to Each Other.