Colag

Colag is a shared room where a team and its AI coding agents build the same product together. The problem was coding agents are made for one person at one keyboard. The plan gets lost in chat, agents overwrite each other's files, and nobody clearly owns the decision to ship. How Colag works is everyone in the room writes one shared plan and design board with requirements, algorithms, diagrams and drawings. You can connect your own coder, such as Cursor or Claude Code. You can also command Colag's agent team. Atlas drafts the plan and the tasks. Ada designs the architecture. Forge builds the backend and Pixel builds the frontend, each in its own Git workspace so they never overwrite each other. Sentinel runs the tests in a sandbox without internet access and reviews the result. The AMD part: for this hackathon we added a runtime that runs these agents on AMD Token Factory, using DeepSeek V4.1 Flash on AMD GPUs. The agents only see Colag's own tools, so permissions stay exactly the same. We built it for a shared team key: it waits and retries when the rate limit is reached, and it checks every structured answer from the model before using it. A person approves the plan and a person merges the tested code. No agent and no model can do either. The code is open source under the MIT license.