harness-sdk

Why we included this project

Building an agent that survives contact with real users means solving problems the notebook demo never shows you, and this SDK wraps the agent loop in the pieces that make that possible: context management, execution limits, and tracing are handled for you rather than left to ad hoc code. It is genuinely model-agnostic, with solid support for Bedrock, Anthropic, OpenAI, Gemini, and local providers like Ollama, and the monorepo ships both Python and TypeScript SDKs so the same agent logic can drive a Python backend or a Node service. Hooks let you intercept every step to validate or redirect behavior, guardrails catch mistakes before a tool runs, and steering handlers give agents a way to self-correct instead of failing silently. Multi-agent patterns, MCP, streaming, and structured output come included, which makes it a reasonable starting point for a team that wants one dependency to cover the common production concerns.