infyrence/infy

Infy is a Python framework for building AI agents that talk to language models, call tools, and follow multi-step plans. It is written from scratch with no required third-party libraries in its core, and it uses an optional compiled Rust extension to speed up common jobs like parsing JSON and comparing text, falling back to plain Python if that extension is not installed. The project focuses on two things. The first is being light and fast: the maker's own tests across a set of real agents show it starting up much faster and using much less memory than other popular agent frameworks, while writing roughly the same amount of code. The second is control: infy includes an optional governance layer that checks every tool call an agent tries to make against a policy before it runs, and writes a record of what happened that cannot be quietly changed afterward. This is meant to make it safer to let an agent take real actions, such as running shell commands, deploying code, spending money, or touching customer data. The governance system can deny certain actions outright, require a human to approve risky ones through a pluggable approval step, and pause a run for hours until a person responds, then resume it once approved. Every decision, whether allowed or denied, gets written to an audit trail. The same governance layer can also wrap agents that were built using other agent frameworks, not just ones built with infy itself. Beyond single tool-calling agents, infy includes a graph based execution engine for building agents with branching logic, loops, and the ability to pause partway through and resume later, with support for saving progress along the way. Both synchronous and asynchronous versions of the model, tool, and graph interfaces are provided. Infy requires Python 3.10 or newer and installs through pip, with optional extras for providers like OpenAI, Anthropic, Gemini, and Ollama. It is currently in an alpha stage and released under the Apache 2.0 license.