Braintrust
A Slack-like workspace where you and a team of AI agents share channels, files and memory.
The question I was chasing
What does it feel like to work with a team of agents, not just prompt one?
Single-agent chat forgets everything between sessions and can't collaborate. I wanted to know whether agents could share context the way a team shares a Slack workspace: channels, files, a memory they all draw on.
Why it exists
Most agent tools are one human talking to one model in a box. The interesting problems start when several agents and a person work in the same shared space over time, on the same projects, with the same files and history in reach.
The constraints
Agents need to pull the right context across projects without blowing the context window. That constraint is the whole design problem: it forces memory and retrieval to the centre, not the edge. Get that wrong and the agents either forget what matters or drown in what doesn't.
The decisions that mattered
Chroma as a vector database, as the backbone rather than a bolt-on: semantic search, persistent memory and retrieval across workspace projects and agents are the core of the thing, not a feature.
Support both OpenAI and Anthropic in the agent builder instead of locking to one provider.
A Slack-like model, with channels and shared context, because it's an instantly familiar mental model for multiple participants collaborating, which is exactly what a team of agents is.
What it is
A workspace and canvas where you and your team of AI agents collaborate on projects: an agent builder (OpenAI and Anthropic), channels, file uploads and shared workspace context, all sitting on Chroma for semantic search and persistent memory. There's a walkthrough above.
Where it landed
A working prototype, and the clearest thing I've built about where multi-agent collaboration is heading. What I'd change next is how agents negotiate who does what. Right now coordination is looser than it should be.
Part of the Rolling Waves work archive.