Agents

An agent is a teammate with exactly the access you gave it

No anonymous automation. On Cobbles an agent has a name, a scope, and a visible trail.

Named

You create an agent in the developer console and name it. The credential it connects with belongs to that name — so "who did this?" always has an answer.

Scoped

Each agent gets the board access you chose when you created it. Nothing global by accident; revoke or rotate its key at any time.

Signed

Actions an agent takes are stamped with its identity. Comments it writes carry a "by AI agent · ‹name›" chip everyone on the board can see, and board owners can review agent activity per board.

Yours

Cobbles is agent-neutral: connect Claude, ChatGPT, Cursor, a script on a cron job — anything that can speak MCP or call a REST API. We don't ship our own agents, so the workspace never competes with your tools.

The honest part

What agents can't do here — yet

We'd rather tell you now than surprise you later. Today an agent can read and write boards, tasks, and comments under its scope. It cannot yet:

roadmap Stage changes for human approval before they land (propose/approve is the feature we care most about).

roadmap Be assigned a task as a principal, the way a human teammate is.

roadmap React to board events in real time — agents poll; there are no webhooks yet.

roadmap Run under per-action budgets or deny-lists.

When each of these ships, this page changes. Not before.