Each agent I build does one job, shows its evidence, and writes nothing without a person approving it. This guide walks through seven of them: what each one does, what it must not do, and the control that keeps it honest.
The rule behind all of them is in AI drafts, people decide.
In this guide
- The deal progression agent: one next step per open deal
- The call recap agent: a transcript becomes a CRM note and a follow-up
- The customer health agent: a rating is only useful with its evidence
- The customer handoff agent: what a new account owner needs on day one
- The data agent: ask the CRM a question in plain language
- The QA agent that checks the other fourteen
- Why my design agents don’t grade their own work
The deal progression agent: one next step per open deal
Deal reviews usually fail in one of two ways. They list every deal and leave no time to act, or they cover three deals and ignore the rest. The deal progression agent I built does the boring part first.
For every open deal, it reports the momentum, what is blocking it and one concrete next step. That is the whole job. It doesn’t send anything and it doesn’t change a record. It writes a draft and waits.
Why one next step
A list of five suggestions is a list nobody follows. One next step per deal gives the seller a decision to make: do it, change it or skip it. If a deal has no obvious next step, that is also a finding.
Why the blocker comes first
Momentum tells you how a deal is moving. The blocker tells you why it isn’t. A deal with good momentum and a named blocker needs a different conversation from a deal with neither.
The control
Like every agent in the set, it drafts and a person approves. Nothing reaches the CRM, an inbox or a customer until someone has read it. The agent can be wrong about a deal it doesn’t have full notes on, and the person reading it usually knows what it can’t see.
The call recap agent: a transcript becomes a CRM note and a follow-up
The work after a call is repetitive: write the note, update the CRM, write the follow-up. It is also the work that gets skipped when the day runs long.
The call recap agent takes a transcript and turns it into a CRM note and a follow-up email. Both are drafts. A person reads them, edits them and approves them before the note is saved or the email is sent.
Why it works as a draft
A transcript has the facts, and the agent’s job is to organize them. The seller’s job is to check what was agreed and add what the transcript couldn’t capture, like tone or a side comment. That split is the 80/20 rule I use for every agent.
What to watch for
- Commitments. Check that every promise in the follow-up was actually made.
- Dates. A date that was only floated should not appear as agreed.
- Names and numbers. Compare each one with the transcript.
- Consent. Only use recordings you are allowed to use.
The control
The agent proposes, and the record changes only after approval. If the note is wrong, it is wrong in a draft, not in the CRM.
The customer health agent: a rating is only useful with its evidence
A red account with no explanation is an alarm with no information. The customer health agent I built gives a rating and shows why.
It rates each account green, yellow or red, cites the evidence behind the rating and suggests next actions. The evidence is the product. The color is just the summary.
Why the citation matters
If a rating points to a quote, an email or a record, a person can check it in seconds. If it doesn’t, the person has to trust it or ignore it, and both are bad outcomes. This is the same rule as every agent in the set: each claim carries its source.
Why next actions
A rating tells you where to look. A next action tells you what to do when you get there. An account that is yellow for a clear reason has a different fix from one that is yellow because nobody has talked to the buyer in a while.
The control
The agent drafts the ratings and a person decides what to do with them. It doesn’t contact a customer or change an account record on its own.
The customer handoff agent: what a new account owner needs on day one
When an account changes owners, the history usually lives in one person’s head. The customer handoff agent writes it down.
When an account moves to a new owner, the agent produces a complete transition report from the records it can read. The goal is that the new owner can pick up the account without calling the old one.
What a good handoff covers
- Who is who: the people at the account and their roles.
- The history in a few lines, not a timeline of every email.
- Open commitments: what has been promised and by when.
- Risks: what could go wrong and what is known about it.
- The first things the new owner should do.
What stays with a person
A report can’t carry a relationship. The new owner should still meet the old one, or the customer, and use the report as the agenda. The agent organizes the record, and the people finish the handoff.
The control
The report is a draft. The outgoing owner reads it, corrects it and approves it before it goes to the new owner.
The data agent: ask the CRM a question in plain language
Most CRM questions go unasked because a report takes too long to build. A data agent lowers that cost.
The data agent answers questions in plain language from the CRM, and queues any change it proposes for approval. You can ask it which deals have no next step or which accounts changed owner this quarter, and read the answer instead of building a report.
Reading is safe, writing is not
Answering a question changes nothing. Changing a record is different, so every proposed change goes into a queue and is shown as a table of before and after. A person approves it before it is applied.
How to check an answer
- Ask for the records behind the number, not just the number.
- Spot-check two or three against the CRM.
- Be careful with questions that depend on how a field is used. A field named “stage” can mean different things to different teams.
The control
The agent reads freely and writes only through approval. That is the line I keep for every agent that touches the CRM.
The QA agent that checks the other fourteen
A set of agents drifts. You fix one, and something that used to work stops working. A QA agent exists to catch that.
In the Sales Agent Intel Suite, a QA agent checks the other fourteen against their own specifications. It is a regression check. When something changes, it tells you whether the agents still behave the way they were written to behave.
What a specification gives you
A written specification says what an agent should do, what it should never do and what a good output looks like. Without one, “it seems fine” is the only test, and that fails quietly.
What a QA agent can and can’t do
- It can check an output against the written rules.
- It can catch an agent that stopped citing its evidence, or one that stopped asking for approval.
- It can’t tell you whether the rules are right. A person still owns that.
The control
The QA agent reports and a person decides what to fix. The suite is built as Claude skills for HubSpot and isn’t released yet. I’m happy to walk through it on request.
Why my design agents don’t grade their own work
An agent that builds a page and then says the page is good has marked its own homework. I split the jobs.
The design team I use to produce my own pages and documents has nine agents behind one command: a strategist, an art director, a copywriter, three builders for web, documents and dashboards, a critic, a QA engineer and a scout that keeps a reference library current.
Three separate jobs
- The builders make the work.
- The critic scores it against a written rubric and sends it back for revision.
- QA measures things that can be measured: contrast, layout at small widths and page weight.
Why the split matters
Taste and measurement are different checks. A critic can say a page feels cluttered. QA can say a text color fails a contrast threshold. Keeping them apart means one doesn’t excuse the other.
The rule underneath
Nothing ships on one agent’s say-so, and a person makes the final call. It is the same principle as the revenue agents: AI drafts, people decide.
The projects page has the short version.
Where to go next
You can see all fifteen agents and how each one is controlled on the AI agents page.
To run something like this yourself, Sales workflow skills for Claude is sixteen skills that draft while you approve, and Claude and HubSpot: connect it safely covers the connection. To have it built for you, see RevLathe.
I’m Seth Levanen. I find the AI shortcuts that keep the quality and drop the busywork. Email me if you want to talk about it.