harness-sdk

harness-sdk is the Strands Agents monorepo for building and running AI agents in Python and TypeScript. The official README describes Strands Agents as an open-source SDK for teams that would otherwise write their own agent loop. It runs in the developer's process rather than forcing a hosted control plane, and it packages the pieces that production agent loops tend to need once a demo becomes real. The repo covers lifecycle controls such as turn limits, token budgets, cancellation, and stop reasons. It also documents tools, structured output, MCP integration, multi-agent patterns, memory, sessions, model portability, streaming, guardrails, tracing, and evals. That combination makes it useful for teams that want a reusable agent foundation instead of rebuilding execution, observability, and safety controls for every project. At review time, GitHub listed the project as strands-agents/harness-sdk, primarily Python, Apache-2.0 licensed, with more than 8,300 stars, more than 1,200 forks, and a recent push on September 24, 2026. The README links to documentation at strandsagents.com, sample projects, a Strands MCP server directory, Discord, PyPI packages, and npm packages. Package badges reference strands-harness, @strands-agents/harness, @strands-agents/cli, strands-agents, and @strands-agents/sdk. The strongest fit is a builder or platform team moving from one-off agent demos to controlled agent execution. A support bot, research assistant, coding workflow, or internal operations agent all need budgets, tool boundaries, session handling, traces, and evaluation hooks. Strands gives teams a source-visible starting point for those concerns while keeping model choice portable. Pricing is recorded as free open-source software for the SDK. Production costs depend on the selected models, provider rates, tool calls, hosting, storage, tracing, and traffic. Teams should measure token usage early, add stop conditions before long-running loops, and test tool calls under realistic failure modes. OpenTools classifies harness-sdk as a developer tool. It supports MCP, but the durable entity is a broad agent SDK and harness, not a single MCP server. Start with a small workflow, add tools one at a time, turn on tracing and evals, and use guardrails before connecting sensitive actions or customer data. For evaluation work, treat Strands as agent infrastructure rather than a finished application. Pick one model provider, one simple tool, and one measurable task. Add token budgets, cancellation, structured outputs, and trace review before expanding to multi-agent patterns or sensitive tools. If the workflow needs MCP, verify the server permissions and the data returned by each tool call. That staged approach lets teams learn the SDK without creating an unbounded autonomous system on the first day.