What is gmas?

gMAS is a Python framework for building systems made up of multiple AI agents that work together, where the roles each agent plays and how they are connected to each other can change while the system is actually running. Instead of coordinating agents through scattered prompt templates and callback functions, gMAS keeps the whole team of agents, their connections, and their state in one structure called a role graph. In this graph, each agent is a node with its own instructions, model, tools, and local memory, and the connections between agents carry work and context along a specific path. A scheduler looks at this graph and figures out the order agents should run in, including which ones can run at the same time. As the system runs, it keeps a record of what each agent produced, how many tokens it used, how long it took, and any errors, and it can change its own plan mid-run through defined rules rather than by improvising, such as stopping early, skipping an agent, rerouting work, or recovering from a failure. The framework does not depend on a specific AI provider. It works with any function that takes a text prompt and returns a text response, so it can connect to OpenAI-compatible services or other model providers depending on how that function is written. It also supports tools such as running shell commands, searching the web, reading files, and connecting to the Model Context Protocol, along with retry logic, timeouts, and budgets to keep runs from running away with cost or time. A more advanced feature lets the system's structure change while it is running: a piece of code called a topology hook watches what happens after each step and can respond by stopping the run, skipping or forcing certain agents, adding or removing connections between agents, or changing which agent gives the final answer. gMAS also supports treating the agent graph as data for graph machine learning, including exporting it in formats usable with PyTorch Geometric for graph neural network research. It requires Python 3.12 or 3.13, installs through pip or uv, and is released under the MIT license. The project is described as an early release aimed at developers and researchers, with the core API tested but some advanced parts still expected to change.