The whole point of this framework is to increase the robustness of agents so that small models (30b class) can function over long time horizons reliably. Other goals include auditability (full agent sessions are stored durably and can be branched/rolled back/restarted from any point, all over-the-wire comms are also durably stored) and reliability (sane fallbacks for common failure cases).
The current iteration (sketch 7 or draft 7 or 8) is specifically a coding agent framework. In the future I'd like to expand the core to handle a variety of tasks.
1. an extension to my shop (I have a shop! I never thought I would have one of these) for water treatment and food storage.
2. an agentic framework that I started on in January - maelstrom. You can find the current code here: https://github.com/zerohumancompany2/maelstrom-code/tree/ske...
The whole point of this framework is to increase the robustness of agents so that small models (30b class) can function over long time horizons reliably. Other goals include auditability (full agent sessions are stored durably and can be branched/rolled back/restarted from any point, all over-the-wire comms are also durably stored) and reliability (sane fallbacks for common failure cases).
The current iteration (sketch 7 or draft 7 or 8) is specifically a coding agent framework. In the future I'd like to expand the core to handle a variety of tasks.