Union is an open-source framework developed by Union Labs aiming to simplify and democratize the creation and deployment of AI Agents. It tackles the complexities of building agents that can reliably perform tasks by providing a structured approach to defining agent capabilities, managing memory, and orchestrating tool usage. At its core, Union focuses on enabling agents to “specialize” – becoming highly proficient in specific domains rather than attempting general intelligence. This specialization is achieved through a unique approach to defining agent behavior using a declarative “Union Schema.”

The central concept is the Union Schema, a YAML-based configuration file that defines everything about an agent: its role, goals, the tools it has access to, and crucially, how it should *use* those tools. Unlike traditional prompting approaches that rely heavily on natural language instructions, the Union Schema uses a more structured and predictable format. This schema dictates the agent’s reasoning process, specifying the steps it should take to achieve a goal, including which tools to call, what inputs to provide, and how to interpret the results. This declarative nature makes agents more reliable, reproducible, and easier to debug. The schema also allows for defining constraints and guardrails, preventing the agent from taking undesirable actions.

Union’s architecture is modular and designed for scalability. It consists of several key components. The `Union Agent` is the runtime environment where the agent executes its tasks based on the defined schema. The `Union Registry` is a central repository for storing and managing Union Schemas, allowing for easy sharing and reuse of agent definitions. `Union Tools` are the external services or APIs that the agent can interact with – these can range from simple calculators to complex databases or web search engines. A key component is the `Union Memory` system, which allows agents to store and retrieve information relevant to their tasks, enabling them to learn and improve over time. Union supports various memory backends, including vector databases like Pinecone and Chroma, as well as simpler options like in-memory storage.

A significant advantage of Union is its focus on observability and control. The framework provides detailed logging and tracing capabilities, allowing developers to understand exactly how an agent is reasoning and making decisions. This is crucial for identifying and fixing errors, as well as for ensuring that the agent is behaving as expected. Furthermore, Union allows for human-in-the-loop intervention, enabling users to review and approve agent actions before they are executed, adding a layer of safety and control. The framework also supports A/B testing of different schemas, allowing developers to optimize agent performance.

The repository provides extensive documentation, examples, and a growing community. It includes pre-built schemas for common tasks, such as web search, data analysis, and code generation, serving as starting points for building custom agents. Union is actively developed and supports integration with popular LLMs like OpenAI's GPT models, Anthropic's Claude, and open-source alternatives. Ultimately, Union aims to lower the barrier to entry for building powerful and reliable AI agents, empowering developers to automate complex tasks and unlock new possibilities with AI.