What our agent is not allowed to do

Everyone is selling what their agent can do. Ask the other question.

You want the leverage. You have watched the demo where it books the meeting, writes the follow-up and updates the record, and some part of you wants that badly, because you have done all three of those things at eleven at night.

Another part of you is picturing the same software writing to the most important person in your industry, in your name, at 2am, in a register you would never use. You cannot un-send that. There is no undo for a first impression made on your behalf.

So you do what careful people do. You try it on something that does not matter, and it stays there. It never gets near the deals that do, which is the whole reason you were interested.

An agent you cannot bound is one you cannot use for anything that matters. The limits are the product.

Here is where ours stops, in the order it matters.

There are three live levels, and you set them per deal rather than once for your whole account. It finds and ranks, and you write and send. It may prepare research on people you are already talking to. It may sort inbound within limits you set. Every one of those is enforced on the server before anything goes out, which is a different thing from being described in a settings screen.

There is a fourth level, and it is the interesting one. It would let the agent make first contact: approach somebody you have never spoken to, in your name. That level is defined in the code and deliberately left unbuilt. Software that cold-approaches strangers in your name is a reputational risk with no upside we can justify, and the day it exists is the day the other three levels stop meaning very much.

The rung for cold contact does not exist to be switched on.

When two members both have agents, the agents may talk to each other, and that is bounded too. They may ask each other exactly four questions, with three exchanges per deal counted across both directions, and only after both people have separately agreed to it. Then it goes back to the humans, whether or not it finished. A negotiation that runs out of turns is a negotiation two people can pick up. A negotiation with no limit is one nobody is supervising.

HubSpot is the fair comparison here, because it is shipping the same category of thing to far more companies than we are. Its Prospecting Agent researches prospects and writes outreach, and its own documentation, updated on 14 August 2026, gives the company using it two settings: review before sending, or send automatically.

Approval there is a toggle, and it nearly has to be. A horizontal platform sold to 299,458 paying companies cannot make one company’s restraint a permanent property of the product, because restraint is exactly the thing its customers disagree about.

Be precise about what that does and does not mean, because the loose version is checkable and wrong. HubSpot does hard-code some restraint: its agent will not send more than three cold emails to one contact within 90 days. So “no guardrails” is false, and an informed reader will say so. The narrow claim, and the only one worth making, is about per-message human approval, and whether it can be switched off.

Here it cannot, and the boundary for a given deal is not a preference at all. It is a sentence you wrote about that deal, in your words, which the matches are then ranked under.

What that adds up to, across a week, is less dramatic than a demo and more useful.

On Monday nothing in her world changed, so the night is skipped entirely and no brief arrives. On Tuesday there is one match, with the reason it surfaced and a warm path. On Wednesday she writes a short note and sends it herself. On Thursday a deal is at terms, both parties in the same region, 12M on the table, and it is suggested that they meet in person, along with what could not be checked: it knows a region, not a city, so both in Europe might mean 3,000km apart. On Friday, nothing. Silence is the common answer, and a product that cannot say nothing will eventually say anything.

Four decisions in a week, and it never once acted without her.

That is the shape of what we mean by the term we use for all of this.

Software that works inside a live deal and stops where judgment starts. It does the reading, ranking and explaining; the person does the writing and the deciding; the line between them is one you set per deal, in your own words.

So when you are evaluating any of these products, including this one, ask what it will not do. Then ask whether that answer is a setting or a property.