Introduces PANDA, a decentralized architecture enabling scalable, fault-tolerant coordination among large numbers of heterogeneous LLM agents with dynamic interaction governance.
PANDA (Planning Agents in a Networked Decentralized Architecture) is a decentralized multi-agent system designed to solve multi-step tasks using large, heterogeneous collectives of LLM agents. It addresses scalability, fault tolerance, flexible orchestration, and governance by allowing agents to self-organize into specialized teams without a central coordinator.
Key capabilities include: Scalable Communication: It decouples collective communication (using gossip protocols) from team communication (using reliable TCP), enabling agents to join multiple teams and load-balance tasks efficiently. Flexible Orchestration: PANDA supports three planning topologies—star, chain, and mesh—allowing teams to adapt their structure to specific task requirements. Fault Tolerance: The system detects infrastructure and orchestration failures, using redundant execution and a swap protocol to replan and replace failed agents, maintaining 100% task completion where other systems fail. Decentralized Governance: It employs a web-of-trust model with depth-bounded reachability, enabling agents to establish trust relationships and govern interactions across different administrative domains.
Evaluated on the HotPotQA benchmark, PANDA scales to thousands of agents, assembles teams in milliseconds, and achieves state-of-the-art accuracy at up to 8x the efficiency of baseline systems.
PANDA is presented as a decentralized architecture for coordinating large populations of heterogeneous LLM agents without relying on a single, monolithic orchestrator. Rather than routing all task decomposition, scheduling, and communication through a central control point, the architecture distributes coordination across agents and orchestration components that can dynamically shape how agents discover one another, exchange capabilities, request work, and respond to failures. Its core emphasis is on flexible orchestration: the interaction topology and control flow are treated as runtime-adaptive rather than fixed in advance, allowing the system to reconfigure agent roles, communication paths, and task allocation in response to workload changes, agent availability, or policy constraints.
The key contribution is the combination of decentralization with dynamic interaction governance. PANDA appears to address a central weakness of many current multi-agent systems: centralized orchestrators can scale poorly, become single points of failure, and struggle to accommodate diverse agent types with different models, tools, latency profiles, and trust boundaries. By making orchestration flexible and governance explicit, the architecture can enforce rules about who may interact with whom, what information may be shared, how requests are escalated, and how the system should recover when agents or links fail. This suggests a design in which scalability and fault tolerance are not bolted on after the fact, but emerge from the coordination substrate itself—through local decision-making, adaptive routing, redundancy, and decentralized state management.
The work matters because it targets one of the main obstacles to deploying LLM agents at production scale: reliability under heterogeneity. Real-world agentic systems are unlikely to consist of a single homogeneous model type; they will mix specialist models, tool-using agents, memory components, human-in-the-loop controls, and policy constraints. A decentralized, governance-aware architecture like PANDA offers a more plausible path toward large, self-healing agent fleets that can maintain throughput and correctness as components fail or workloads shift. In short, it reframes multi-agent coordination as a systems problem—where orchestration, fault tolerance, and governance must be designed together—rather than merely a prompting or workflow problem.