GPTSwarm
Type: full-code · Vendor: metauto-ai · Language: Python · License: MIT · Status: active · Status in practice: experimental · First released: 2024
Represent LLM-based agents and their connections as optimizable graphs whose edges and prompts are tuned automatically toward better swarm performance.
Description. GPTSwarm is a graph-based framework for LLM-based agents. It builds agents from computational graphs and connects them into swarms whose inter-agent edges can be pruned or created at run time. The framework includes optimization algorithms that tune prompts and the graph itself toward better task performance. It is written in Python and released under the MIT license; the underlying work was published as 'Language Agents as Optimizable Graphs' and accepted at ICML 2024.
Agent loop shape. Agents are constructed as graphs of operations and connected into a swarm; an optimizer adjusts the inter-agent edges and prompts against an evaluation signal, pruning or creating connections so the collaboration structure is learned rather than fixed.
Primary use cases
- building LLM agents from computational graphs
- optimizing inter-agent connections at run time
- automatically tuning agent prompts and swarm topology
Key concepts
- Agent as graph → automatic-workflow-search (docs) — Each agent is built as a computational graph of operation nodes; GPTSwarm lets you construct LLM-based agents from these graphs rather than as monolithic prompts.
- Swarm → dynamic-topology-routing (docs) — A composite graph that connects multiple agent graphs together, whose inter-agent edges are the optimizable surface tuned at run time.
- swarm.optimizer → automatic-workflow-search (docs) — The module of optimization algorithms that tunes prompts and graph edges via reinforcement learning and prompt optimization to raise agent and swarm performance.
Patterns this full-code implements —
- ·Dynamic Topology Routing
Inter-agent connections in the agent graph are optimized by pruning or creating edges at run time, reshaping the collaboration topology instead of using a fixed structure.
- ·Automatic Workflow Search
GPTSwarm treats the agent swarm as an optimizable graph and ships optimization algorithms that tune prompts and the graph structure to improve agent and swarm performance.
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.
Alternatives & relatives
- full-code · framework
AgentVerse also studies multi-agent self-organization and emergent collaboration, but configures agent groups declaratively rather than learning the inter-agent graph through an optimizer.
- full-code · framework
AutoGen
AutoGen composes multi-agent systems as configurable conversation graphs; GPTSwarm differs by making the graph and prompts an optimization target tuned automatically toward a reward.
Listed as alternative by (1)
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: partial