Agency

Agency is a fast, minimal, open-source Python framework for developers building systems where autonomous agents communicate and act. It strips away the orchestration bloat, offering a clean, Pythonic API for defining agents, actions, and messaging patterns. Aimed at AI/ML engineers, researchers, and hobbyists, Agency gives you full control over agent logic without locking you into a heavyweight ecosystem. You can prototype quickly on one machine, then scale out to distributed setups when you're ready. Agency's core strength is its two communication spaces: a local space for in-process agent messaging and an AMQP space for distributed, broker-backed communication. Both asynchronous messaging and synchronous request-response patterns are supported, so you can model fire-and-forget tasks or tightly coupled agent calls. The framework also provides essential utilities: agent callbacks (on_start, on_stop) for lifecycle hooks, audit logging via agency.logger, resource management via agency.resources, schema definitions via agency.schema, and an agent queue via agency.queue. The official docs include an example application walkthrough, API documentation, and a GitHub repository under the MIT license. The source code is transparent—you can read exactly how messaging, callbacks, and spaces are implemented. Agency doesn't bundle LLM integrations, a GUI, or managed infrastructure; it's a focused builder's kit. If you need enterprise-grade orchestration with built-in LLM connectors, you'll want to pair Agency with other tools or look elsewhere. But if you prefer rolling your own agent architecture in Python with nothing but the essentials, Agency is a solid starting point. Recent news about AI agents and security highlights the importance of understanding agent behavior, but Agency itself is a development framework, not a security tool. Its minimal design lets you implement your own safety and monitoring layers.