Multi-agent workspaces

One Cosmos can host many scoped sub-brains. Building a research agent plus a coder plus a reviewer with OpenCode, Cline, or Cursor? Give each agent its own folder so they never poison each other's memory, while still living inside a single Cosmos.

How scoped sub-brains work

Each agent gets its own /Agents/<name> folder, seeded with three files:

• prompt.md — the agent's system prompt.
• context.md — background fed to the agent every turn.
• rules.md — what the agent can and can't do.

The external AI connects its MCP scoped to that folder, so each agent reads and writes only its own memory. A researcher's findings never leak into the reviewer's brain unless you choose to share a folder.

Create an agent

1. Navigate to /Agents in the folder tree.
2. Open the ... menu and choose + New agent.
3. Name it — for example, Reviewer.
4. Edit prompt.md (the system prompt), context.md (background fed every turn), and rules.md (what it can and can't do).
5. Connect your AI tool's MCP scoped to /Agents/Reviewer.

Repeat for each agent — for example a researcher and a coder — so every role has its own isolated sub-brain.

Shared vs isolated brains

Agents can share one folder to build a single team brain, or stay fully isolated so their memories never mix. Pick per role:

• Isolated — best when agents have conflicting context (a critic vs. an author).
• Shared — best when agents collaborate on the same task and should learn from each other.

How do I stop agents from poisoning each other's memory?

Scope each agent's MCP connection to its own /Agents/<name> folder. Because recall is folder-scoped, one agent can't read another's memory unless they share a folder.

Which AI tools work with multi-agent workspaces?

Any MCP-capable tool — OpenCode, Cline, Cursor, and the others in Connect your AI tools (MCP) — can connect scoped to a single agent folder.