xyz-ai-lab/axisagentic

AxisAgentic is a software framework for running AI agents that need to complete long, multi-step tasks, such as searching the web across many steps or working through a complicated research question. It manages the back and forth between the AI model and the tools it uses, keeps track of the growing conversation context so it does not run out of room, and can recover and retry when something goes wrong partway through a run. A key part of the project is that it records everything the agent does as it works. Every step of a task is written to a detailed, permanent log called a trace, which captures exactly what the AI model could see at each point. Because that record is complete and unaltered, it can later be replayed to reconstruct a run, used to score how well the agent performed against a benchmark, or exported as training data to teach another AI model to perform similarly. The project intentionally keeps the running of agents and the collection of training data on the same underlying record, so what a model learns from matches what actually happened during execution. The repository includes two example uses built with this framework, one for general web search and one for a more complex multi-part search task. It also references XYZ-Aquila, a search system built on top of AxisAgentic and benchmarked against several other agentic search systems, though the framework itself is meant to support other kinds of agents beyond search, including general purpose or coding focused ones. Note that this repository does not include any AI model weights, only the runtime and tooling around them. Getting started requires Python 3.12 or newer and access to an OpenAI compatible model endpoint, along with running a setup script and configuring provider details before a first task can be executed. The project is released under the Apache 2.0 license, which permits free use, modification, and commercial use. This project is aimed at AI researchers and engineers building or evaluating long running autonomous agents, particularly teams that want to collect high quality training data directly from real agent runs.