Braintrust
A Slack-like workspace where agents share memory and find what they need.
Working with a team of agents
I built Braintrust when the idea of working in a shared workspace with collaborating agents still felt new to me. I wanted to see if I could work with multiple agents in one place, with each of them able to interact, share context and collaborate on the same project.
A shared workspace
The proof of concept worked. I could create agents with their own personas and configure them to work together. The workspace used a familiar, Slack-like layout, with channels, file uploads and shared project context.
The agent builder supported both OpenAI and Anthropic models. Shared memory let agents find related notes across projects and carry context between sessions, so they could draw on the same material as they worked.
Where the complexity remained
Getting that working taught me a lot, but I never felt Braintrust was ready to open source or release. The complexity of coordinating the agents and managing their context was still something I needed to work through. It was a working proof of concept, but the harness wasn't production-ready.
What I learned from it
Braintrust gave me a way to get further into the APIs and some of the more technical questions around how agents were evolving. It was useful to push those boundaries and learn what it took to have several agents working in a shared space.
Seeing similar ideas emerge in tools from the frontier AI labs has felt validating. It was a direction I wanted to explore, and building the prototype helped me understand both the possibilities and the work still needed to make it ready for other people.
Part of the Rolling Waves work archive.