Cumora: agents as first-class teammates

What's this about?

yetone open-sourced Cumora, cross-platform team chat in which AI agents are full participants alongside humans — the same roster, the same direct messages, the same group channels, the same Kanban board and calendar. The interface looks like Slack, but most of the colleagues are AI. Agents have names, personas and memory; they claim work, coordinate with each other, can send and receive real email, and can speak without being addressed — "I noticed last week's issue is still unresolved." The default team includes Atlas (researcher), Bram (engineer), Iris, Nova and Saga. Two runtimes: Cumora Cloud, where agents run in hosted Kubernetes pods on OpenAI's Responses API, or BYOA — bring your own agent — running Claude Code or Codex on your own machine via `npx cumora agent`.

What exactly changes when an agent has a persona and memory?

Today's assistants are stateless and reactive: you ask, they answer, the session ends and they forget. A persona plus memory turns that into something continuous — the agent remembers what it did last week, has a defined role and scope, and knows what it's responsible for. The critical difference is initiative: it can raise something without being asked. That flips the interaction from tool-use to colleague-like, which is the whole design bet — and also why "you almost can't tell who is human and who is AI in the interface" is either the feature or the warning, depending on your view.

Why is this worth noticing?

1. The BYOA option is the thoughtful part. You can point it at Claude Code or Codex running locally, so the agents use tools you already trust rather than a vendor's hosted runtime.
2. Making chat the substrate rather than an IDE is a real design position — it puts coordination between agents, and between agents and people, at the center instead of code editing.
3. It lands the same week as Databricks' argument that enterprise agents finally work because organizations captured the context in people's heads. A shared team channel with persistent memory is one concrete way to capture exactly that.