Topological Coherence for Self-Evolving Multi-Agent Systems

Multi-agent systems can divide a difficult task among specialized agents, but coordination becomes fragile when responsibilities, communication, tools, and memory are designed separately. This paper introduces ToCoMAS, a framework that derives these parts from a shared task topology. Its central requirement is topological coherence: task dependencies should remain consistent with which agents own each region, which handoffs are required, and which memories can be accessed—even as the system evolves.

What the paper does

The authors start with a directed task graph whose nodes represent task operations and whose edges represent dependencies. ToCoMAS grounds every task node in a public graph of tool interfaces and resources. This grounding considers capability coverage, input-output compatibility, side effects, and continuity between tool providers.
Rather than assigning one agent to every task node, the system groups compatible nodes into responsibility regions. Compatibility can reflect shared tools, compatible inputs and outputs, dependency adjacency, and semantic similarity. These regions are then contracted into a quotient topology that preserves dependencies between regions while hiding dependencies internal to one region. A merge is allowed only if the resulting structure remains executable and acyclic.
Each region is assigned to an agent whose available tools cover the region’s requirements. The assignment does not have to be one-to-one: a single agent may own several compatible regions, while regions with different operational requirements remain separate. If no current agent has the necessary capabilities, the system can add a specialization to the agent pool.
This topology-grounded organization differs from simply giving agents textual roles. An agent’s responsibility is defined by the task region and the tools needed to execute it, rather than by a prefabricated label.

How collaboration and memory are coupled

Once responsibility regions and their owners are established, ToCoMAS derives the collaboration structure. A dependency crossing from one region to another creates a required handoff whenever different agents own the two regions. The runtime communication graph may also include additional profile-conditioned links based on agent similarity, query relevance, and complementary tool interfaces.
Handoffs are boundary-aware. A downstream region receives the artifact or portion of an artifact that its task requires, rather than automatically receiving all of the upstream agent’s information. The routing policy can also leave a message unsent or request verification when appropriate.
Memory follows the same ownership boundaries. Each record is associated with its producing agent, task region, and provenance. A region first receives an admissible view of memory according to ownership and sharing permissions; semantic retrieval can rank or expand only within that permitted view. In other words, a relevant memory cannot bypass a boundary merely because it matches the current query.
The framework also addresses ownership changes during evolution. When a region moves to another agent, its private memory is transported with the region, retaining provenance while updating access authority. This is intended to prevent both orphaned experience and accidental global sharing.

Self-evolution with structural checks

ToCoMAS treats the responsibility map, agent ownership, collaboration graph, and memory policy as a coupled structural projection of the task-conditioned system. Execution traces and accumulated experience guide proposals that may revise some of these policies.
A proposal that changes ownership must also reconcile the handoffs and memory permissions affected by the new boundaries. The system then applies a structural validity screen: every task node must have a capable owner, required cross-region dependencies must have compatible handoffs, and memory visibility must respect ownership. Only valid candidates that improve evaluated reward over their parent are retained for later tasks.
The objective combines execution quality with measured resource cost. The paper therefore frames evolution not as unrestricted search over agent configurations, but as reward-based selection constrained by whether the resulting organization can still execute the task coherently.
This focus on controlled decisions and routing has some conceptual overlap with REFLEX’s separation of bounded control choices from open-ended generation, although ToCoMAS addresses multi-agent structure, responsibility, collaboration, and memory rather than selective model invocation.

What results stand out—and what remains open

The authors evaluate ToCoMAS on BBEH for reasoning, WorkBench for stateful tool use, SWE-Bench-Verified for repository repair, and CoMemBench for executable multi-agent workflows with typed handoffs and paired clean and polluted instances. They compare fixed and evolving variants with direct-call, prespecified, automatically designed, and self-evolving baselines.
According to the paper, ToCoMAS improves task success over baselines across backbones on the first three benchmark families and on CoMemBench. On CoMemBench, it also improves over the self-evolving baseline in verified progress, accepted handoffs, and memory isolation. The benchmark’s metrics distinguish terminal success from verified task-node progress, mandatory handoff acceptance, and isolation under memory pollution.
These results support the authors’ claim that jointly enforcing structural consistency can provide benefits beyond independently optimizing agents or communication graphs. They do not, however, establish that every component must change together at every evolution step: the method explicitly allows proposals to modify only a subset of policies, provided the resulting relationships remain compatible.
The material describes the architecture, validity constraints, benchmarks, and headline findings, but does not provide detailed numerical results, ablations, or failure analyses here. The reported evidence therefore supports the value of topology-aware organization in the tested settings, while leaving the magnitude of individual design contributions and behavior on other task distributions less specified.