Very Large-Scale Multi-Agent Simulation in AgentScope

Reading guide

Problem and motivation

Large language model (LLM) agents enable interactive and realistic simulations without extensive manual rule programming. However, scaling multi-agent simulations to realistic demographic sizes reveals three critical bottlenecks:

1. Scalability and efficiency limits: Centralized execution and Python's Global Interpreter Lock (GIL) hinder agent execution. Existing actor frameworks (e.g., Ray) spin up a new process per actor, wasting resources when handling millions of lightweight, I/O-bound agents.
2. Unsatisfied population distributions and agent diversity: Agents with simple, identical background prompts behave homogeneously, failing to represent realistic demographic distributions (e.g., across education, age, or occupation).
3. Difficult management: Orchestrating, launching, monitoring, and terminating hundreds of thousands or millions of agents distributed across clusters requires substantial operational overhead.

Key ideas

• Actor-based distributed architecture: Agents compute independently as actors. The platform supports both one-to-one (agent-per-process, avoiding GIL for compute-heavy tasks) and many-to-one (agents share processes, maximizing CPU core time-sharing during I/O waits) mappings.
• Zero-refactor distributed conversion (to_dist): Users author centralized workflows, which are transparently converted to distributed setups using local proxies and asynchronous message placeholders.
• Environment as an agent: Environments are modeled as specialized agents supporting state maintenance, high concurrency access via RPC, bi-directional triggers (via listeners), and nested multi-layer structures for group-level simulations.
• Configurable heterogeneous population pipeline: Users define demographic distributions in configuration files, and an LLM meta-prompt pipeline generates diverse, individualized agent personas.
• Agent-Manager: A web-based graphical interface and persistent remote server network to deploy, monitor, and clean up agents across cluster nodes.

Method

Distributed Execution & Automatic Conversion

In AgentScope, independent agents execute concurrently. When a centralized pipeline is tagged with .to_dist(), two stages occur:

1. Distributed agents spawn across remote workers; lightweight proxy objects remain on the central orchestrator.
2. Proxies immediately return message placeholders upon receiving a call. Downstream agents only block when they inspect placeholders that point to pending computation from preceding agents.

Simulation Dynamics

The platform evaluates multi-agent reasoning on the classic "guess $\frac{2}{3}$ of the average" game. Each agent $i$ submits $x_i\in [0,100]$. The winning target is: $$x^{\ast }=\frac{2}{3}\cdot \bar{x}=\frac{2}{3}\left(\frac{1}{N}\sum _{i=1}^N{x}_i\right)$$ When iterated under common knowledge of rationality, the unique Nash equilibrium converges to $0$. In alternative experiments testing generalization beyond prior training data, the target rule was adjusted to: $$x^{\ast }=5+\frac{1}{2}\bar{x}$$ Setting $x^{\ast }=\bar{x}$ yields a non-zero Nash equilibrium: $$x^{\ast }=5+\frac{1}{2}{x}^{\ast }⟹x^{\ast }=10$$

Results

• Scalability: A simulation of 1,000,000 agents completed in 12 minutes using 4 nodes (each with 8 $\times$ A100-80G GPUs, 64 CPU cores, 1 TB RAM) using Llama3-8B with basic prompts, and in 85 minutes with Chain-of-Thought (CoT) prompting (which expanded token counts over 150-fold). In a dummy compute benchmark removing LLM inference latency, 1M agents ran in 40 seconds, compared to 8.6 hours via standard Python asynchronous I/O and ~12 days via serial execution.
• Horizontal scaling: With a fixed 10,000 agents per device, overall runtime remained constant as cluster size increased from 1 to 4 nodes.
• Prompting & reasoning depth: Under basic prompting, most models reported around 50 (though MistralAI-8x7B averaged 36.63 and MistralAI-8x22B averaged 31.69). Adding CoT instructions pushed reports toward 0 (Qwen2-72B winning numbers dropped from 33.70 to 12.76, and multi-round round 5 averages dropped from 25.16 to 2.02).
• Demographics & roles: Higher assigned education levels systematically produced lower, more game-theoretically optimal guesses (Ph.D. agents reported lower numbers than elementary school agents). Occupations like game theory professors and economists reported the lowest numbers.
• Variants: When the equilibrium was shifted to 10 ($5+\frac{1}{2}\bar{x}$), agents iteratively converged around 10 over multiple rounds.

Limitations and open questions

• Role-playing inconsistency in edge cases: When prompted to act as a 7-year-old child, agents still executed multi-step arithmetic and recursive reasoning that contradicted the persona.
• Simple arithmetic errors: Agents occasionally made basic calculation slips (especially involving fractions or decimals), even when their strategic reasoning logic was sound.
• Sensitivity to training priors: Shifting the target ratio from $\frac{2}{3}$ to $\frac{51}{100}$ degraded performance because models leaned on memorized game theory patterns; performance recovered only when prompts explicitly stated that the game was a variant of "guess $\frac{2}{3}$ of the average".