strands-agents/harness-sdk

Repository Scope and Entry Points

Strands Agents is a monorepo for model-driven AI-agent SDKs in Python and TypeScript, plus an MCP documentation server, developer CLI, documentation site, and AWS-backed integration-test infrastructure. Public SDK entry points are strands-py/src/strands/__init__.py and strands-ts/src/index.ts (with Node-specific exports in strands-ts/src/index.node.ts). The central invocation APIs are strands-py/src/strands/agent/agent.py and strands-ts/src/agent/agent.ts.

Agent Execution Core

Each SDK centers on an Agent that owns prompt/conversation state, a model, tools, and lifecycle controls. Agent invocation drives an iterative model/tool loop, producing results and streamed events. Python orchestration is split between strands-py/src/strands/agent/agent.py, strands-py/src/strands/event_loop/event_loop.py, and strands-py/src/strands/event_loop/streaming.py; TypeScript consolidates core invocation in strands-ts/src/agent/agent.ts with tool dispatch in strands-ts/src/agent/tool-caller.ts.

Model Provider Boundary

A shared model abstraction isolates agent logic from LLM transports, schemas, streaming, retries, and provider-specific error behavior. Python implementations include Bedrock, Anthropic, Gemini, OpenAI, and local/aggregator providers under strands-py/src/strands/models/model.py, strands-py/src/strands/models/bedrock.py, and strands-py/src/strands/models/openai.py. TypeScript counterparts are strands-ts/src/models/model.ts, strands-ts/src/models/bedrock.ts, and strands-ts/src/models/openai/model.ts. Amazon Bedrock is the default; external services include AWS, Anthropic, OpenAI, Google, and others through adapters.

Tools, MCP, and Execution Safety

Models select registered tool schemas; executors validate and invoke them sequentially or concurrently, returning results to the loop. Core Python paths are strands-py/src/strands/tools/registry.py, strands-py/src/strands/tools/executors/concurrent.py, and strands-py/src/strands/tools/mcp/mcp_client.py; TypeScript paths are strands-ts/src/registry/tool-registry.ts, strands-ts/src/tools/executors/concurrent.ts, and strands-ts/src/mcp/client.ts. Sandboxes separate tool execution from host environments using local POSIX, Docker, or SSH implementations: strands-py/src/strands/sandbox/docker.py and strands-ts/src/sandbox/docker.ts.

Extensibility, Control, and State

Hooks, middleware, plugins, and interventions provide interception boundaries around agent/model/tool stages for logging, authorization, human approval, steering, and cancellation. Key paths are strands-py/src/strands/_middleware/stages.py, strands-py/src/strands/interventions/registry.py, strands-ts/src/middleware/stages.ts, and strands-ts/src/interventions/handler.ts. Conversation/context managers constrain token growth, while session, storage, and memory layers persist or retrieve state; e.g., strands-py/src/strands/agent/conversation_manager/conversation_manager.py, strands-py/src/strands/memory/memory_manager.py, strands-ts/src/session/session-manager.ts, and strands-ts/src/storage/s3-storage.ts. Durable integrations include filesystem and Amazon S3; Bedrock Knowledge Bases are vended memory stores.

Multi-Agent, Observability, and Streaming

Agents can compose as tools or execute through graph, swarm, and Agent-to-Agent (A2A) protocols. Primary implementations are strands-py/src/strands/multiagent/graph.py, strands-py/src/strands/multiagent/a2a/server.py, strands-ts/src/multiagent/graph.ts, and strands-ts/src/a2a/express-server.ts. Telemetry traces and metrics agent decisions and operations through strands-py/src/strands/telemetry/tracer.py and strands-ts/src/telemetry/tracer.ts. Python additionally exposes experimental bidirectional text/audio agents for Gemini Live, Nova Sonic, and OpenAI Realtime under strands-py/src/strands/experimental/bidi/agent/agent.py.

Supporting Products and Delivery

strands-mcp/src/strands_mcp_server/__main__.py starts a Python MCP server that fetches, processes, indexes, caches, and serves project documentation. strandly/src/cli.ts is the workspace developer CLI. The Astro documentation application is rooted at site/src/pages/index.astro, content is defined in site/src/content.config.ts, and custom rendering/routing resides in site/src/route-middleware.ts. GitHub Actions package, test, release, and deploy the SDKs and site; AWS CDK integration-test resources are defined by test-infra/bin/test-infra.ts and test-infra/lib/stacks/test-infra-stack.ts.