ATC: Agent Trajectory Control for Role-Contract Enforcement in Black-Box Multi-Agent Systems
| Source: arXiv AI
Tags: Multi-Agent Systems, Agent Trajectory Control, Role Drift, LLM Agents, Agent Governance
Agent Trajectory Control (ATC) enforces explicit role contracts in multi-agent LLM systems by checking each message before it enters the shared conversation, reducing role drift without access to model internals.
Details
Multi-agent LLM systems split work across planners, retrievers, verifiers and executors, but over many turns an agent can start doing another role's job. The authors call this role drift. ATC makes each role's authority an explicit contract. Before a message is admitted, it is scored for out-of-role deviation, leakage into other roles, impermissible speech acts and logical contract breaches; a constrained regeneration follows, and the revised message is re-checked. Because enforcement acts on messages rather than the model, it works with black-box agents. The authors report better task performance and role consistency than prompting, memory, workflow-control and activation-editing baselines, with the largest gains in role-confusion stress tests. This is relevant for anyone running agent pipelines where an executor should never act without approval.