Strands Agents

A Strands Agent runs a model-driven loop: the model reads the conversation, decides on tool calls, the SDK executes them and feeds results back, and the loop ends on a stop reason, a turn or token limit, or cancellation. Hooks fire at each lifecycle event so code can log, validate or redirect a step. The newer Strands harness is a thin factory (create_harness() / createHarness()) that returns a pre-configured Agent with built-in tools, sub-agents, memory and context management.

Multi-agent work has four documented shapes. With agents as tools, an orchestrator agent calls specialist agents like any tool. Swarm hands a task between agents that choose their own successor, sharing a history of handoffs. Graph follows developer-defined nodes and edges, where conditional edges pick the path and cycles are allowed; nodes can themselves be swarms. A Workflow is a fixed DAG of tasks run as one tool. A2A connects agents across processes or frameworks.

State is kept per agent in its conversation and state objects and persisted through session managers (in Python SnapshotSessionManager for new single-agent sessions, FileSessionManager or S3SessionManager for Graph and Swarm). For multi-agent systems only the orchestrator holds a session manager. The harness also keeps long-term memory of distilled facts across sessions by default.

The SDK has an interrupt mechanism: a hook callback or tool raises an interrupt, the agent loop stops and returns the pending interrupts, and the caller supplies responses to resume from the same point. The HumanInTheLoop intervention handler builds on this to let a person approve, edit or reject a tool call, with modes for CLI, web or custom UIs and optional session-level trust. Interrupts also work inside Swarm and Graph runs, and the harness adds presets, natural-language rules and Cedar policies for gating tool calls.

pip install strands-agents

from strands import Agent
from strands.multiagent import Swarm

# Default model: Claude on Amazon Bedrock (us-west-2), so AWS credentials are needed.
researcher = Agent(name="researcher", system_prompt="Gather key facts, then hand off to the writer.")
writer = Agent(name="writer", system_prompt="Write a short, clear answer from the facts you receive.")

swarm = Swarm(
    [researcher, writer],
    entry_point=researcher,
    max_handoffs=6,
    execution_timeout=300.0,
)
result = swarm("Explain what a B-tree index is in three sentences.")
print(result.status, [node.node_id for node in result.node_history])

• The Python SDK needs Python 3.10+; the TypeScript SDK (npm install @strands-agents/sdk) needs Node.js 22+.
• Amazon Bedrock is the default provider (Claude Sonnet in us-west-2): set AWS_BEARER_TOKEN_BEDROCK or normal AWS credentials, and enable model access in the Bedrock console. Other providers need extras such as pip install 'strands-agents[anthropic]'.
• Ready-made tools like calculator live in the separate strands-agents-tools package.
• The assembled harness is a different package (pip install strands-harness) from the SDK (strands-agents).
• Agents placed inside a Graph or Swarm must not have their own session manager; Python raises ValueError.
• README links under /docs/user-guide/concepts/... point to paths that moved; use the current docs navigation.

It was consolidated into a monorepo with the TypeScript SDK, docs and MCP server and now lives at strands-agents/harness-sdk. The announcement says PyPI and npm package names and install commands did not change.

MCP (client) and A2A (client and server) are built into both SDKs. AG-UI works through a partner-maintained, Python-only package according to the Strands docs.

The SDK is Apache-2.0 licensed and runs in your process with no hosted control plane. The default model provider, Amazon Bedrock, is billed by AWS; other providers bill separately.

The SDK (strands-agents) gives you the agent loop, tools, providers and multi-agent primitives. Strands harness (strands-harness) is a factory on top that returns a pre-configured agent with built-in tools, memory and context management.

The docs suggest Graph for fixed flows with conditional branches or loops, Swarm for open-ended collaboration through handoffs, Workflow for a repeatable DAG run as one tool, and agents as tools for simple delegation.