Agents
An agent is a job, its rules and a record, with its own key. You describe the job in plain words. Layrcake writes the plan, tests it on real data and sends nothing until a person says yes.
What an agent is
An agent is a plan you can read, a key of its own, and a record of everything it did. Its plan says who it looks for, what it sends, who approves, how much a run may spend and where it runs. Its key carries the same rules as any Layrcake key: your do-not-contact list, suppression, caps and approvals hold on every call, wherever the agent runs.
| Kind | What it does |
|---|---|
| Outbound | Finds companies and the right person at each, checks their address, drafts the emails and builds the campaign. You approve the launch. |
| Watcher | Watches companies and reports what it finds. It never sends anything. |
| Reply handler COMING | Sorts replies and drafts answers for your yes. Its test isn't available yet, and an agent goes live only after a passing test. |
The Agents page lists your agents with their status (Draft, On, Paused), this month's credits, people emailed, replies, and any launch waiting on you.
The builder
Choose Create an agent on the Agents page. The builder is a chat on one side and the plan on the other, in four steps.
- Describe. Say what the agent should do. One sentence is enough, for example “Find heads of RevOps at UK SaaS companies hiring three or more sales roles, and email an intro and a follow-up.” Or start from a recipe.
- Plan. Read the plan it wrote. Change any line with Change, or say what to change in the chat.
- Test run. Run it on real data. Nothing is sent.
- Go live. Put it live. It gets its own key.
The builder keeps the rules for you. Ask it to send without your approval and it says no. Name a data provider and it sends you to Connections instead of adding it as a tool. Anything not live yet is tagged COMING on the plan and can't be switched on. Anything else you ask for that has no row of its own is kept as an Also line, in your words.
Recipes
| Recipe | Status |
|---|---|
| List to campaign | Runs today |
| Signal-triggered outbound | Runs today |
| LinkedIn and email | The email part runs today. Checking each person and recording every LinkedIn touch is coming. |
| Meeting prep | Reading your calendar needs a calendar tool you bring. Writing what happened to HubSpot is coming. |
| Reply handler | Coming |
| Win back from the CRM | Coming |
| Inbound reply | Coming |
| Customer expansion | Coming |
| CRM clean-up | Coming |
| Weekly report | Coming |
Test run
A test runs the plan on real data and stops before anything is sent. No email goes out, no campaign is created, nothing is bought.
Before it starts you see its price: about how many credits it should use, split into data and model. Confirm and it runs. The credits come from your balance as each step settles, and the test is held to a hard ceiling a little above its quote, so it can't run away.
An outbound test runs these steps, and shows each one as it goes:
- Find up to 10 companies that match Who.
- Pick the person at each, by the roles in the plan.
- Find their addresses, and check each one works.
- Leave out anyone on your do-not-contact list.
- Draft the intro and the day-4 follow-up.
- Stop at the launch preview.
Then it checks the result. An outbound test has eight checks:
| Check | What it asks |
|---|---|
| List fit | Does each person match the brief? A person counts only when their company passed the check against Who and their title matches one of the plan's roles. |
| Addresses | Were the addresses checked, and do they work? |
| Do-not-contact | Was everyone you may not contact left out? |
| Approvals | Did it stop and wait for a yes, with nothing sent? |
| Cost | What did it cost per contact? It reports; it only blocks if it fails. |
| Copy | Are the emails short, specific and free of claims the plan can't back up? |
| Sending address | Can your workspace send? Without a sending address the real run can't build the campaign. The test links to the Sending page. |
| Stops cleanly | Did it stop at the launch preview? |
A failed check has a Fix button. It sends a sentence to the chat, the plan changes, and you test again. A test that found nobody says which step came back empty and offers one fix, usually to widen who it looks for.
Testing again reuses, free, every paid step that hasn't changed: the companies already found aren't paid for twice. If you change the plan after a test, that test is out of date and you test again before going live.
Test results hold personal data. They are dropped 30 days after the agent goes live or is retired, or 60 days after the test for a draft that never goes live.
Go live
Go live opens only after a passing test of the plan as it is now. It needs someone who can manage keys in the workspace.
- Read the last check: the agent's name, its first run, who approves before anything is sent, and what it costs. An agent that runs in Layrcake also shows its thinking cost per run and its monthly ceiling.
- Read the price line. Your first agents are included. A paid one shows its monthly price, and is charged to your card only when you confirm.
- Choose Go live. The agent turns on and gets its own key.
Approvals
A launch always waits for a person's yes. For any agent that emails people cold, the plan's Asks row is locked to “You approve every launch”. No chat message, plan edit or import can change it.
- An outbound agent builds the campaign as a draft and stops at the launch preview. The launch comes to Approvals in the console. Nothing is sent before a named person approves it.
- A watcher never sends anything.
- An agent's key, or a run in Layrcake, can ask for an approval but can never give one. Only a signed-in person can.
See Human confirmations for how approvals work on every key.
Next
- Where agents run: in Layrcake or in Claude Code, Codex and n8n, schedules and Run now.
- Budgets, versions and control: the per-run budget, the monthly cap, versions, pausing and retiring.