AI Agents

Agents are automation you describe in plain English. Paste an input — a customer email, a question, a campaign brief — and the agent does the legwork: it reads your data, plans the steps, and proposes any changes. Reads run immediately; every write waits for your approval and is logged. That one-line promise is the whole design:

"Agents run any process you describe: paste the input, review what they propose, approve once. Reads are free; every write waits for your approval and is logged."

Agents are available on every plan.

Creating an agent

Press New agent and give it:

• a Name
• What does this agent do? — the instructions, in plain English (for example: "Analyse our P&L for the last year and tell me the top 3 things driving profit and the biggest risk.")
• Actions — what it's allowed to touch, picked from three groups:

Planning (safe, run immediately)

Action	What it does
Xero contacts	Search contacts; propose contact updates (approved by you)
Xero P&L report	Read-only: fetch profit & loss for analysis
Stripe payouts	Read-only: list payouts (website / Pitchup)
Bookings lookup	Read-only: search bookings by guest

Changes (proposed, need your approval)

Action	What it does
Create Xero draft invoices	Proposes DRAFT invoices — never sent until approved in Xero
Send email reply	Proposes a single email from your connected mailbox
Send bulk / marketing email	Proposes an email to a list — every recipient shown before approval, capped at 50

Guided manual steps

Action	What it does
Zoho contact (guided)	Guided manual step with the values to paste
Client details sheet (guided)	Opens your Google Sheet with values to paste

You're in control

"The agent can only use the actions you enable. Write actions always show you exactly what they will do before anything happens." An agent without write actions literally cannot change anything.

Running an agent

1. Press Run on the agent's card.
2. Paste your input — "Paste the customer email, a question, a campaign brief…" — and press Run agent.
3. The agent plans, reads what it needs, and returns:
   • a plain-English report
   • a Proposed changes list (if it wants to write anything)
   • Show what the agent did (N steps) — the full trace of every call it made

Approving or rejecting

Each proposal shows exactly what will happen — the email address and subject and body; the invoice's lines and total; the field-by-field contact diff. Then:

• Approve & apply — the run status becomes Applied
• Reject — the run status becomes Rejected

Some items are guided steps rather than automatic writes — after approval they show a Mark done button next to the values to paste, so the run ends with everything either applied, done, failed or skipped. Every run is kept (Recent runs, last 15) and can be reopened read-only at any time.

Safety rails

Rail	How it works
Approval is mandatory	v1 agents always require approval — it cannot be turned off
Drafts, not sends	Xero invoices are always DRAFT — you approve inside Xero
Bulk cap	Bulk emails capped at 50 recipients per run, recipients shown up-front
Connected tools only	Writes need the matching integration (Xero, mailbox) connected first
Full trace	Every tool call is logged: tool, arguments, result

Example agents

• "Answer this guest email." Actions: bookings lookup + send email reply. Paste the email; review the draft (it already knows the booking and your tone preferences); approve.
• "Month-end invoicing." Actions: Stripe payouts + create Xero draft invoices. Paste "invoice all payouts in October"; review each proposal; approve.
• "P&L health check." Actions: Xero P&L report. Paste "last 12 months"; read the report — no approvals needed because it only reads.