GPTSwarm is an open-source research framework developed by the metauto-ai group that represents LLM-based agents as computational graphs. In this graph model, nodes implement functions that process multimodal data or query LLMs, while edges define information flow between operations. Agent graphs can be recursively composed into larger graphs, enabling hierarchical multi-agent collaboration. The framework's defining capability is automatic graph optimization. It supports two levels of optimization: node-level prompt refinement (improving the instructions given to individual agents) and edge-level connectivity optimization (restructuring how agents communicate and pass information). These optimizations can be driven by reinforcement learning or prompting-based techniques, allowing swarms to self-improve on a given task without human intervention. GPTSwarm was published as an academic paper (Zhuge et al., 2024) titled 'GPTSwarm: Language Agents as Optimizable Graphs' and was accepted to ICML 2024 as an Oral presentation, placing it in the top 1.5% of submissions (144 out of 9,473 papers). The code is publicly available on GitHub under the metauto-ai organization. The framework targets AI researchers and developers who want to build scalable, adaptive multi-agent pipelines. Practical use cases demonstrated include software code generation and review, multi-step research automation, and process automation. The swarm approach has been shown to be robust to adversarial perturbations within the swarm and outperforms individual agents on several benchmarks. As of mid-2025, the project's most recent documented milestone was its ICML 2024 presentation; active maintenance status in 2026 was not confirmed in available sources. Key features: - Graph-based agent representation where nodes are LLM calls or functions and edges are information flows - Automatic node-level prompt optimization via reinforcement learning or prompting - Automatic edge-level graph connectivity optimization to restructure agent communication - Recursive graph composition to build hierarchical multi-agent societies - Self-improving swarms that optimize without human intervention during task execution - Adversarial robustness: swarms are resilient to perturbations within the agent graph - Open-source Python library installable via pip (github.com/metauto-ai/GPTSwarm) - ICML 2024 Oral presentation (top 1.5% of submissions)
Not public (open-source research framework; no commercial pricing tier found)