The Harness: How an AI Agent Gets Trusted With a Sale

Most AI in finance stops at the chat box. You ask a question, it answers, you decide what to believe. That works until the question is "what is my practice worth, and what do I do about it," and the answer is going to be negotiated over with real money on the table.

So the interesting question is not whether an agent can help with a succession or a sale. It can. The question is what keeps it from helping too much.

The third act in a play we have been writing

As far as I can tell, this is the third thing ValuationMCP has done first in this category, and each one set up the next.

First came the MCP-native app: a valuation you can run from inside Claude or ChatGPT, instead of a website you have to visit. Then came agent to agent: your agent can talk to a buyer's agent, with a person approving each step, instead of everyone emailing spreadsheets. I wrote about both on LinkedIn; the links are in Sources.

Each step gave agents more reach. Reach is exactly the problem. The harness is the third act because it is what makes the first two safe to use on real work.

What a harness is, without the jargon

Think of a smart new hire on day one. You would let them read anything in the file. You would welcome their suggestions. You would not hand them the company card on the first morning.

And there are a few things you would never let them do alone, however capable. Send a message to a client in your name. Change what a client pays. Sign the NDA. Decide that the work is finished.

That is a harness. It is not the intelligence. It is the set of rules around the intelligence: what it may look at, what it may suggest, what needs your sign-off, and what it can never do.

The models will keep getting smarter. That is not the part I worry about. A smarter model with no rules around it is just a faster way to send the wrong thing to the wrong person. The rules are what turn a capable model into something you would let work on a sale.

Four lanes for every action

Every tool an agent loop can call at ValuationMCP is registered in one of four lanes. Try sorting a few.

Read runs freely. Propose turns into a card on your desk, and nothing happens until you act on it. Approve runs only after your explicit click. Never is refused and written to the log; there is no code path behind it to get lucky with.

The lanes are not a policy document sitting next to the software. They are how the tools are registered, and the diagram below reads them from the same list.

Pulls from your systems are worth a sentence of their own. The agent never starts one. It proposes the pull, you start it, and only the totals you approve are saved, each with its source. A buyer's agent never pulls from a target's systems at all. It asks, through a data request, under an NDA.

The model writes the words. The engine owns the numbers.

This is the rule that matters most when the numbers decide who gets paid.

The agent can put the issues in your data in order, phrase the next question, and draft a plan. It does not state, compute, round or restate a figure. Every figure comes from the deterministic valuation engine, shown as a range with its data coverage. Where the model writes text, a sentence with a digit in it is dropped before you see it. Figures appear only as facts the app copies from the engine.

It also cannot grade its own homework. A move is marked verified only by a saved valuation run, never by the agent saying it worked. And drafts stay drafts: the agent prepares the request, a person sends it.

What a loop looks like

An agent run is not a free-roaming conversation. It is a loop with a fixed shape, a step log, and caps on how far it can go. Step through the one that checks your data before every valuation.

Two details are worth noticing.

The log only grows. Each step is written once, with a key so it cannot be recorded twice. A run that is interrupted resumes from its last step, and a click that lands twice does not do the work twice. It also means anyone can read back exactly what happened, in order.

And the loop stops and waits for you at the right moments. Two sources disagree on revenue? It does not pick one. It asks. Waiting is a feature.

Where it runs today

The same harness runs every loop, with the same four lanes and the same rules.

For an owner, that is the loop that checks the data before any number is computed, and loops that work the operating margin, the client book, billing and growth. Ledger, the hosted agent, works an assigned move on its schedule and leaves every change as a card.

For a deal, it is two loops, one for each side. The buyer's loop finds what is missing and proposes requests. The target's loop drafts answers from its own record. They exchange messages over A2A, nothing is drafted or shared until an NDA is in place, and an incoming message only ever becomes a card for a person to read, never an action.

Why this is the real moat

Anyone can wire a model to a data source. In succession and M&A, the failure that ends the product is one confidently wrong number, or one message sent that should not have been.

A harness is how you make that failure structurally unlikely, instead of promising to be careful. I think this category will end up sorting vendors by it.

It is also, frankly, the boring part. Nobody demos a permission lane. Nobody screenshots a step log. But an owner deciding whether to let software near a sale does not need a clever answer; they need to know what the software is not allowed to do, and that the list is enforced by something other than good intentions. A buyer needs the same thing from the other side of the table. So does the M&A advisor sitting between them, who has to explain to both where every figure came from.

MCP-native, then A2A, then the harness. I am not claiming the story is over, only that these were the first three chapters and the order was not an accident.

What it does not do

It is not an appraisal, a fairness opinion, or investment, legal or tax advice. It does not replace your attorney, your CPA or the buyer's diligence. It is a way to let an agent do the legwork while the decisions, and the sending, stay with people.

If you want to see the parts, the harness page walks through them, and the agent-to-agent page shows the deal loops. If you want your firm's own number, run a valuation in Claude and watch where it stops and asks.