When to Start Communicating: Adaptive Stigmergy Gates Improve Multi-Agent RL Training Dynamics

Shared communication channels are often treated as an unconditional good in multi-agent reinforcement learning (MARL): giving agents access to messages, shared memory, or stigmergic traces should improve coordination. Yet communication can also destabilize learning when it is available before it is informative, inducing spurious correlations that interfere with credit assignment and representation learning. We present an Adaptive Stigmergy Engine that controls when agents may communicate through a shared pheromone field. The  engine uses a simple gate evaluated mid-episode to activate the field only under population-level distress signals (population decline, energy crisis, or delivery inequality).                                                                                              

We evaluate the approach in a JAX multi-agent foraging simulation trained with PPO and evolutionary pressure (birth, death, and reproduction). Across 10 seeds in a clean end-to-end run (10M training steps) and a 20-experiment development arc, adaptive deployment outperforms both ALWAYS_ON and ALWAYS_OFF baselines, improving hard-task deliveries by +34% and +19% respectively while yielding the lowest population variance. Cross-domain validation with LLM agent collectives (481 runs across two models and three task domains) confirms the principle: on adversarial tasks with planted decoy bugs, communication exposure increases false consensus capture monotonically (0% at 0 shared rounds to 100% at 3 shared rounds), demonstrating that communication costs depend on information quality in both RL and LLM settings.

Website: Atlaso Research