By Steve, S (@sstvee11), building Bloome
## 1. When Agents Start Working Together
The first time several useful agents share the same workspace, it feels like leverage.
The second time, the coordination problems start to show.
When multiple agents work in the same public environment, the hard part is no longer making one agent useful. It is making useful agents work together.
That is why we started building an Agent Collaboration Protocol.
We ran into this while building Bloome, an agent-native workspace where people and AI teammates share the same conversation. A user might ask:
> Can you three review this launch plan? One of you check product risk, one check engineering risk, and one check go-to-market. Please give me a final recommendation.
Each agent can understand the request. Each agent can produce useful work. But useful individual work does not automatically become useful teamwork.
Without a collaboration protocol, the failure is not that agents are "bad". The failure is that they are acting from separate local decisions inside a shared public workspace.
Three patterns show up quickly:
• agents duplicate the same part of the work;
• agents answer from stale context after another agent has moved the task forward;
• the user becomes the manager who has to reassign, correct, and merge the work.
That pushed us to ask a more basic question:
> What is the smallest task that exposes the same coordination failure?
We used counting as the minimal benchmark:
> Count from 1 to 20, one agent at a time, without duplicates, and stop at 20.
Without the protocol, agents see the same room but act from separate local guesses.
Generated by Thread Navigator
Press ⌘ + S to quick-export