Heterogeneous Agent Cohorts for Safe Open-Ended Exploration with Runtime Constraint Memory

This paper addresses the fundamental conflict in autonomous agents between creative exploration and runtime safety. Rather than forcing a single model to balance these competing objectives, the author proposes a heterogeneous cohort consisting of three specialized roles: a Disrupter (which generates high-entropy, unconventional proposals), a Validator (which enforces hard runtime safety checks at the tool gateway), and a Broker (which retrieves out-of-domain knowledge to bridge semantic gaps).

To manage the exploration-safety trade-off, the system employs two primary innovations. First, it uses Contrastive Novelty Retrieval (CNR), which selects external knowledge that is highly relevant to the task goal but semantically distinct from the current dialogue history, preventing the agents from falling into homophilous traps. Second, it implements a persistent memory mechanism called Scars. When an agent triggers a safety violation, the system uses Monte Carlo Tree Search (MCTS) to compile the failure trace into a compact, signed constraint patch. These patches are cached and inherited by future agent cohorts, effectively turning repeated failures into permanent, low-cost runtime constraints that prevent the model from proposing known unsafe actions.

In a spatial-semantic sandbox, the cohort successfully reached remote targets that traditional multi-agent debate frameworks failed to achieve. The Validator ensured that no safety breaches were executed, while the Scars mechanism reduced token consumption by 15.1% by eliminating redundant validation checks. Furthermore, the author introduced a Communication Allocation Score (CAS) system, which dynamically manages bandwidth based on role-specific rewards, reducing total token costs by 55.9% under resource-constrained conditions.