Build production-ready agent controls without writing the loop from scratch
Imagine a scenario where you can instantly see what decisions an agent made, where it stopped, and how many tokens it used. This repository makes it possible to implement that level of insight with just a few lines of code. In short, it is an open-source toolkit that bundles all the features needed to run and control AI agents.
Options for finished agents and custom loops
This repository provides tools at two levels. First, calling the create_harness() function immediately generates an agent optimized for memory, session, and context management. This setup includes the basic configuration needed to move ideas into production, allowing you to start quickly. On the other hand, if the defaults are insufficient or specific logic is required, you can directly configure the agent loop, tools, model providers, and memory using the Python and TypeScript SDKs. This structure allows you to handle everything from initial prototyping to building complex custom agents within a single ecosystem.
Built-in features for control and visibility
Tracing functionality, which tracks every decision process while the agent is running, is enabled by default. This allows you to record what happens at each step and insert hooks to verify or adjust direction. Additionally, guardrail features catch errors before they execute, and steering handlers guide the agent to correct its own mistakes. These features prevent agents from failing silently and contribute to deriving reliable results.
Flexibility independent of models and clouds
Major model providers such as Amazon Bedrock, Anthropic, OpenAI, and Gemini are treated with first-class support, and you can also connect Ollama or custom providers. Because you are not dependent on a specific cloud or model, you can replace the backend as your scale grows while keeping your existing code intact. MCP (Model Context Protocol) servers, streaming, multi-agent patterns, and structured output are also built in, making integration with external tools and real-time response handling easy.
Things to check before applying
The Python SDK requires version 3.10 or higher, and the TypeScript SDK requires an ESM project environment with Node.js 22 or higher. When using TypeScript, Amazon Bedrock is used by default, so you must set up AWS credentials or explicitly pass a different model. This project is licensed under the Apache License 2.0, allowing commercial use. If you plan to build complex agent systems, it is important to select the components that fit your workflow through the provided documentation and examples.
