Jobs an operator can hand to an agent.

Prepare the work, then wait for the person who owns it.

Most of the time in a serious desk is preparation: read the packet, pull the relevant record, draft the reply, fill the fields. That is the part an agent can take. The part a person still has to own — sending it, spending against it, changing the system of record — is a step you mark. The run stops there, shows the reviewer what it prepared and what it used, and continues only on that decision.

The wait is part of the run, so it is still there after a restart. The approval names who decided, and it sits in the same record as the draft. A later reader can see the preparation and the human decision as one sequence.

WHAT THAT IS FOR

A desk that moves on the reading and the drafting, and stops where a person has to sign the outcome.

Brief from the corpus already on the appliance.

An operator asks for a briefing from material the organization already holds: a case file, a procedure set, the last cycle of reporting. The agent reads what that agent is allowed to read, on the appliance, and drafts from those passages. The documents stay where they were. The draft is something a person edits, and the passages it used are still attached.

This is the same grounding Substrate gives a person who asks a question, put to work as a step inside an agent. The corpus is scoped. An agent allowed to read one collection does not wander into another because the question was broad.

WHAT THAT IS FOR

A first draft a specialist can correct. Sources that never had to leave the building to produce it.

Act in a system you already operate.

An agent that only drafts is half a desk. The useful ones also file the ticket, post the status, or call the service where the work already lives. On Agency that call is a description you wrote: the address, the payload, the credential the run may use. The description is the boundary. The run can make the calls on that list, and the record shows that it did.

Credentials are there for the run to use. People operating the agent do not read them back out. A new system is a new description, reviewed and published with the version, so an agent gains a tool the way it gains any other step — on purpose, in a version you can name.

WHAT THAT IS FOR

Work that lands in the system of record. A list, written down, of what the agent was allowed to touch.

Publish a version your applications can keep calling.

An agent that other systems depend on needs a stable address and a version you can point to. Publishing freezes the design. Deploying points the live address at that published version. Applications keep calling the same place. When you publish the next one, you choose when the address moves.

A draft is where operators and developers work. Live traffic calls a published version. If the version you just deployed is wrong, you point the address back at an earlier published one. The previous design is still there, because publishing kept it.

WHAT THAT IS FOR

A live agent with a name and a version. A way back that does not require rebuilding it.

Open yesterday's run and read it in order.

A run is a record, not a chat that disappears. It keeps the input, each step, the model output, the passages used, the tool calls, the policy decision on the step, and the person who approved when a person had to. A reviewer can follow it while it is happening. Someone else can open it the next morning and read the same sequence.

That record is what you show when someone asks what the agent did on a particular matter. Publishing, deploying, and changing who may approve are written down beside it, with who did each of those things. The log is built so a change to it is evident.

WHAT THAT IS FOR

An explanation you can hand to your own oversight. The run and the decision in one place.