Swarms

The building block is Agent: a model plus tools, a system prompt and a loop count (max_loops, or "auto" to let the agent decide when it is done). Model calls go through a LiteLLM-based manager, so provider choice is a model_name string.

Multi-agent behaviour comes from a catalog of structure classes rather than one core model: pipelines (SequentialWorkflow), fan-out (ConcurrentWorkflow), DAGs (GraphWorkflow), director-worker hierarchies (HierarchicalSwarm), bidding group chat, mixture-of-agents, routers and more. SwarmRouter puts these behind one call, and SocialAlgorithms lets you write the communication order as a plain Python function.

State is a per-agent Conversation history held in process. Setting persistent_memory=True writes turns through to a MEMORY.md file keyed by agent_name, reloads it on the next run, and compresses older context when the window fills. No shared memory store across agents is documented beyond passing outputs along the chosen structure.

The documented hook is interactive=True (or mode="interactive") on an Agent, which the FAQ describes as pausing for user input between loops; HierarchicalSwarm(interactive=True) only turns on a live terminal dashboard. No per-tool approval gate or checkpoint-and-resume flow is documented, and the v13 changelog says leftover human-in-the-loop code was removed. Searched docs for approve, interrupt, human and interactive.

pip3 install -U swarms

from swarms import Agent, HierarchicalSwarm, SequentialWorkflow

researcher = Agent(agent_name="Researcher", agent_description="Collects facts on a topic.",
                   system_prompt="List the key facts, each with a source.", model_name="gpt-5.4", max_loops=1)
writer = Agent(agent_name="Writer", agent_description="Turns notes into a short brief.",
               system_prompt="Write a 150-word brief from the notes.", model_name="gpt-5.4", max_loops=1)

# Fixed order: the researcher's output becomes the writer's input
pipeline = SequentialWorkflow(agents=[researcher, writer])
print(pipeline.run("Heat pumps in cold climates"))

# Same agents under a director that plans and assigns orders
team = HierarchicalSwarm(name="Brief-Team", description="Research then write",
                         agents=[researcher, writer], max_loops=1)
print(team.run("Compare heat pumps and gas boilers for a small office"))

• Requires Python 3.10 or higher.
• Set a provider key such as OPENAI_API_KEY or ANTHROPIC_API_KEY in the environment; WORKSPACE_DIR controls where agents write files.
• OpenTelemetry tracing is on by default and records inputs and outputs; set SWARMS_TELEMETRY_ON=false before the process starts to disable it.
• persistent_memory defaults to False; reuse the same agent_name to resume a memory file.
• v13 removed the SwarmRouter speaker-function API and renamed get_all_agent_names to return_all_agent_names; check the changelog when upgrading.

Yes, both ways. Agents can call tools from MCP servers through mcp_url or mcp_urls, and MCPDeployer serves agents, swarms or functions as authenticated MCP servers. A2A and AG-UI support were not found in official sources.

The framework is Apache-2.0 and free to use; you pay your model provider. The same team runs a hosted Swarms API with free and paid subscription tiers plus usage-based billing.

The framework is Python (3.10+). The ecosystem page lists a separate Rust framework and API clients for Swarms Cloud in TypeScript, Go and Java.

It does not force one. It ships sequential, concurrent, DAG, director-worker, group chat, mixture-of-agents and router structures, and SwarmRouter selects among them.

Only as a coding aid: the repository ships a CLAUDE.md guide and the docs publish an llms.txt so assistants such as Claude Code or Cursor can write Swarms code. Swarms does not run or plug into Claude Code as a harness.