AI Agents
Overview
An agent is a goal, a set of things it knows, and a set of limits. It writes each email when it is needed, from what has happened in that conversation so far, rather than from a sequence typed in advance.
It is not a new way of sending mail. An agent's messages go out through the same verified sender as a campaign, with the same tracking, the same unsubscribe header and the same suppression list. What it adds is that the next message is written for the person actually in front of it.
You will find them under Email & Agents → Agents.
Where to start
Creating an agent offers a gallery of starting points — book a call, keep warm, win back, or something of your own. Picking one fills in a goal, a tone and a set of limits you can then change.
The hardest part of setting one of these up is knowing what a good goal sounds like, and a blank box teaches nobody that. Start from a template even if you change every word of it.
What an agent needs
A goal
What it is trying to achieve, in your own words. "Get them to book a twenty-minute call with one of our closers, be direct about why the call is worth their time, and give them the booking link rather than asking when suits" is a goal. "Nurture the lead" is not.
A voice
A job title and a tone. The name it writes as is the display name on the address it sends from — it is not asked for separately, because two sources for one fact drift, and an agent that introduces itself as one person and signs off as another is worse than one with no name at all.
A knowledge base
The single biggest difference between an agent that writes well and one that invents things is what it was given to read. Add your product one-pager, your price list, objection handling, a couple of case studies — as a web page, an uploaded file, or pasted text.
Two things worth knowing:
- It is read once, when you add it. The text is stored. A page quietly rewritten in March would otherwise change what the agent says in April with nothing recording that it did. Use Read it again when you want the current version.
- It may only state what is in there. The agent is told this, and then the code checks it. Anything it cannot support from the knowledge base, it does not claim.
Example emails
Optional, and the fastest way to fix an agent that sounds like a machine. Paste two or three emails your team actually sent that you would be happy to receive.
These teach voice, not facts. The agent is told to copy how they are written and none of their content, because the facts in them belong to a different customer.
An audience
Who it works on. A saved lead list, leads carrying certain tags, leads from one source, leads coming up for renewal within a window, or everyone with an email address.
Renewals are a window rather than a saved list on purpose — the people renewing in thirty days are different people next week.
The limits
These are the part that decides whether a real person gets a real email, so they are settings rather than instructions, and they are enforced in code.
- Maximum messages — never more than this many to one person, ever.
- Minimum days between messages — the gap it must leave.
- Daily limit — the most it may send in a day, across everybody.
- What happens when somebody replies — either hand the conversation to a person immediately, or keep talking and hand over when it matters. The second is usually what people want: a "not right now, try me in Q3" does not need to interrupt anyone, and a conversation turning to price does.
- Send without asking — off until you have read a few drafts. See below.
Drafts and approval
With Send without asking off, every message the agent writes waits for you. The approval queue shows the draft, who it is to, and the agent's own reasoning — what it knows about the person and why this message now.
Read a handful. When the drafts stop surprising you, switch auto-send on.
Turning it on does not remove the limits above. It removes the pause.
What a conversation does
Each person the agent works is a run, and a run ends in one of a few ways:
| State | Meaning |
|---|---|
| In progress / Waiting to write | Still going |
| They replied | Somebody wrote back |
| Goal met | The agent decided it had achieved what it was for |
| Stopped | Nothing more worth saying, or handed to a person |
| Out of messages | It reached the maximum and stopped |
An agent is allowed to decide there is nothing useful left to say and stop. A message sent to fill a quota is worse than silence, so that answer is honoured rather than overridden.
Where the model runs
Two options, and the only difference is whose AWS account and whose bill:
- Your own AWS account. Straightline uses the role you already granted it. Your data never leaves your account for inference, and the tokens land on your AWS bill next to your sending charges.
- Straightline's account. For organizations that have not connected one.
Use Check it can reach the model on the agent page before switching it on. Model access in AWS is granted per account and per region from the Bedrock console, so a perfectly good role still gets refused until somebody has ticked that box — far better to find out there than at the moment a message was due.
If something looks wrong
- Nothing is happening. Check the agent is switched on, and that its audience actually matches anybody. The agent page says when its next pass is due.
- It stopped on its own. Agents switch themselves off when the model becomes unreachable, rather than failing once an hour into a log nobody reads. The reason is shown on the page.
- A draft says something wrong. The draft records which sources it relied on. Almost always the fix is in the knowledge base rather than in the goal.
Tips
- Give it less to read than you think, but make all of it true. A model handed a marketing deck full of superlatives writes emails full of superlatives.
- Two or three example emails beat a paragraph describing your tone.
- Leave hand over when it matters on for the first few weeks. Handing over early costs a colleague two minutes; handing over late costs the deal.
- Read the reasoning on the drafts, not just the drafts. It tells you what the agent believes about the person, which is where a bad email starts.