An agentic workspace you can actually leave running

One job, from the message to the record

Everything in the product is a step on the same line. This is what happens between typing a sentence and being able to prove what came of it.

1. You write a message in the inbox

   In a channel or a direct conversation, you describe the job and mention an agent. Mention two and the product works out whether they run in parallel or hand work to each other in sequence.

2. An agent picks it up with its own brain

   A named worker with a job title, written instructions, assigned skills and a brain of its own, the AI tool it thinks with. It can only see the channels it is a member of, and it did not choose which ones those are.

3. It reaches a real system through a connector

   Not with your password: with a credential the workspace holds, encrypted and bound to that workspace, exposed to the agent as named tools. Reads are open; writes follow a per-connection policy an admin sets, and a new connection asks a person first.

4. A write waits for a person

   A connector write on the default policy waits in Approvals and, once approved, runs exactly once. An admin can mark a connector autonomous or refuse its writes, and that change is recorded; a delete waits for a person under every policy. Merges, knowledge changes and finished work from a reviewed-tier agent stop the same way: the agent posts what it wants to do, a workspace admin approves or rejects it with a comment, and the decision does something — it does not merely flip a status.

5. The audit log takes the record and chains it

   Each entry carries the previous entry’s SHA-256 hash. Rewriting one row invalidates every row after it, and a verification pass names the exact sequence number where the chain breaks.