11 · Working with the agents
What you will have done: asked an agent to do revenue work in chat, and understood exactly where it stops.
The one rule
An agent may draft. An agent may never send.
There is no tool, anywhere, that puts a letter on the wire without a person approving it. This is not a setting you could turn off by accident; the tool simply does not exist.
flowchart LR
A["You ask in chat"] --> B["The agent works"]
B --> C["A draft, a list, a note"]
C --> D["Approvals or the Inbox"]
D -->|"you approve"| E["It happens"]
D -->|"you do not"| F["It does not"]
Who does what
Four agents build a campaign’s list, and each answers a different question.
| Agent | Question it answers |
|---|---|
| Artemis | Which companies are worth writing to? |
| Athena | What is true about this company? |
| Themis | Is this company worth the letter, and why? |
| Ariadne | Who at this company should be written to? |
Above them, the Outreach Manager runs the campaign: it decides what happens next and when.
The walkthrough
1. Ask for something in chat
Plain language. The agent has tools for the same routes the screens use, so anything it does appears on the screens exactly as if you had done it.
Useful asks:
- “Find twenty British knitwear makers with no US distributor.”
- “What came back on the US distribution campaign this week?”
- “Read the reply from Aldermoor Leather and tell me what it says.”
- “Draft an answer to Jonathan Church.”
- “Who at Bexley & Crane should we be writing to?”
2. Know what comes back to you
Some things happen straight away. Others become something waiting for your decision.
| You ask for | What happens |
|---|---|
| A list of companies | Built, and lands on the campaign |
| Research on a company | Done, and recorded |
| A person found at a company | Found, and added |
| A letter written | A draft, waiting in the Inbox |
| A letter sent | Refused. No agent tool sends. |
| A label corrected | Done, and the new label’s effects run |
| A campaign paused | Done |
3. Give it a task rather than a question
An agent given a goal does better than an agent given an instruction. “Find me companies like Aldermoor Leather but in knitwear” beats “search for knitwear companies”, because the first says what good looks like.
4. Check its work on the screens
Everything an agent does shows on the ordinary screens: companies on Leads, drafts in the Inbox, letters in Approvals. There is no separate agent view to learn, and no hidden state.
What always needs a person
- A first letter to somebody new. Approvals.
- A drafted answer. The Inbox.
- Un-pausing a domain the warden paused. Journey 8.
- Anything that spends money outside the configured vendors.
A worked example
Meridian Supply ask: “Read the Aldermoor Leather reply and draft an answer.”
The agent reads the thread, sees the reply already labelled positive, and writes a draft citing two approved claims. It comes back with what it wrote and where to find it.
They then ask: “Good — send it.”
The agent cannot. It says so, and points at the draft in the Inbox. They open it, read it, and press Approve & send themselves.
That is the whole design in one exchange.
What to do if
The agent says it cannot send. Correct. Approve the draft yourself in the Inbox.
It built a list of companies you did not want. The brief is what it works from. Sharpen the brief on the campaign’s Setup and ask again — the brief is the instruction, the chat message is only the trigger.
It cites a claim you have never approved. It cannot. If a draft cites something not on the campaign’s list, it is refused where it is written. If you believe you have seen otherwise, that is worth reporting.
It did something you did not expect. Everything is on the screens and on the person’s timeline. The timeline is the record, and it settles disagreements.
Confirm it worked: whatever you asked for is visible on the ordinary screens — a company on Leads, a draft in the Inbox, a letter in Approvals.