Can multi-agent teams automatically remove their weakest members?

DyLAN (Dynamic LLM-Agent Network) introduces a systematic mechanism for multi-agent team optimization that addresses three properties simultaneously: task agnosticism, efficiency, and automatic team composition.

The core mechanism is the Agent Importance Score, computed through a three-step procedure:

1. Propagation — each agent rates its predecessors on their solution quality
2. Aggregation — for each agent, ratings from successors are compiled to quantify its contribution
3. Selection — after summing ratings across all time steps, top-performing agents are retained and low-performing agents deactivated

This creates a dynamic interaction architecture: agents viewed as nodes in a network exchange messages as edges across time steps. An LLM-empowered ranker ranks agents at inference time and deactivates low-performing ones for subsequent rounds, while an early-stopping mechanism prevents unnecessary iterations.

The insight connects to multiple threads in multi-agent reasoning:

Since Why do multi-agent LLM systems converge without real debate?, DyLAN's contribution scoring provides a partial solution — agents that merely agree without adding information would receive low importance scores and get deactivated. This prevents the noise-amplification problem documented in When does debate actually improve reasoning accuracy?.

The approach contrasts with Can extreme task decomposition enable reliable execution at million-step scale? (MAKER), which uses static decomposition with voting. DyLAN dynamically prunes the agent network during execution — a more adaptive but less parallelizable strategy. The trade-off maps onto How should we balance parallel versus sequential compute at test time?: static decomposition enables parallelism while dynamic selection enables adaptation.

The Agent Importance Score also provides a concrete implementation of the "contribution-based routing" that Can AI systems detect when they've genuinely reached agreement? advocates — but generalized beyond agreement detection to overall contribution quantification.

AgentVerse four-stage dynamic group adjustment (from Arxiv/Agents Multi): AgentVerse extends the dynamic team composition principle with a four-stage group problem-solving process that mirrors human group dynamics: (1) Expert Recruitment — dynamically adjusting team composition based on current problem-solving progress; (2) Collaborative Decision-Making — recruited agents discuss and formulate strategies until consensus; (3) Action Execution — agents interact with the environment to execute agreed actions; (4) Evaluation — comparing current state to desired goal, with feedback reward looping back to stage 1 for team re-composition. Unlike DyLAN's contribution scoring which prunes within a fixed network, AgentVerse's recruitment stage can introduce new agent profiles not in the original team. The evaluation-to-recruitment feedback loop enables adaptive team evolution over the course of problem-solving — the team that finishes may differ substantially from the team that started.

MasRouter's four-decision MASR framework (from Arxiv/Routers): MasRouter formalizes multi-agent system routing as four simultaneous decisions: collaboration topology, agent count, role allocation, and per-agent LLM selection. This reveals that DyLAN's contribution-based agent selection addresses only runtime optimization within an already-constructed network. MasRouter constructs the network itself — choosing topology, roles, and LLM assignments from scratch via a cascaded variational-probabilistic-multinomial controller. The two approaches are complementary: MasRouter for initial construction (design-time routing), DyLAN for runtime adaptation (inference-time pruning). Composing them would create a system that starts with an optimal network configuration AND adapts it during execution. See What decisions must multi-agent routing systems optimize simultaneously?.

Source: Agents