Moving the first agent took a week. The second took an afternoon.

Dear Negotiation Explorer,

Week eight of building the NegoAI platform in the open.

In May I showed you four agents that prepare a negotiation together. Kraljic is the buyer expert. Kahneman is the behavioral expert. Deepak is the negotiation expert. The Compiler turns what the three produce into one report. All four lived on Cassidy, a no-code tool on somebody else's machine.

This week they moved onto mine. Deepak first, then Kraljic, then the other three. By the end of the week the whole thing was running on my own server in Germany, the one I told you about in July, buying side and selling side. Brama, my partner in Claude Code, did the moving. I decided what to keep. I read every result.

Deepak is big. His instructions are 237,000 characters. His knowledge base, the negotiation theory I put into him, is another 197,000. On Cassidy all of that was in front of him every time he ran.

My server does not work like that. An agent has one short file that is always in front of him, and the rest sits in drawers he opens when he needs them. So the question was which lines Deepak needs in front of him every single time. The rest can stay in a drawer.

We went through the three files line by line and gave every block of text a place. Brama wrote a program that refuses to build the new Deepak until all 572 blocks have a place, or a written reason for leaving one out. Not a word was summarized or rewritten. The file that is always in front of him came out at 31,000 characters.

Then I had to find out whether he was still Deepak.

Before touching anything, we ran the old Deepak twice on the same case, with the same words, in two separate chats. The second answer was a fifth shorter than the first. It kept the same 48 headings and the same four scenarios for what the other side might do, with the same odds. It gave all four different names.

So the old Deepak does not agree with himself. That difference is my measuring stick. If the new Deepak differs from the old one by less than that, it is the same agent.

Then the new Deepak ran the case five times. The first offers, on an $18.5 million contract, came out at 12.5, 12.75, 12.75, 12.75 and 13.9 million dollars. The target was 13.9. The old one had opened at 12.75 and 13.0. Same agent, same case, and the number you walk into the room with is different every time.

He kept losing the structure. Deepak's preparation always has eleven sections, in a fixed order. The new Deepak kept them on some runs and lost them on others. His instructions said "sections 1 to 11", and the names of those sections sat in a separate file. When he opened that file, the sections were there. When he did not, they were gone. So we put the whole structure, every heading, into the file that is always in front of him. After that he kept it, eight runs out of eight.

Most of his knowledge base never reached him. The server hands an agent a file only up to 100,000 characters. The part of Deepak's knowledge about people, the psychology, was 117,000. So for eleven runs Deepak got the first 1,500 characters and a note saying the rest could not be delivered. And all our records said the knowledge base was in place. We found out by asking him to quote one sentence from it, without looking anything up. He could not. Now the knowledge base is three files instead of one, not a word changed. Ask him again and he quotes it back: Many negotiators have been taught a simple rule: "Never make the first offer."

Somebody else has to check. Every version of the new Deepak was read by an agent that had nothing to do with the move. It compared each run against both runs of the old Deepak, on the nine things a Deepak preparation has to deliver. Every verdict had to carry the line it was judging, and the line number. By the second review the new Deepak matched the old one on all nine, or sat inside the difference the old one has with himself.

On my server every run says "cost included", because it runs on a subscription. A platform will not. It will pay per call. So Deepak ran the same case fourteen times on a model that costs a tenth of the one he normally uses. If we paid per call, one preparation is about twenty cents on the expensive model and about two cents on the cheap one. The checking agent could not tell the two apart on judgment. On facts, the cheap one got something wrong twice in fourteen runs, the expensive one once in nine. Deepak stays on the expensive one. Every agent we move gets the same test.

Kraljic took an afternoon, because by then we had a procedure. 374 blocks placed, and it worked the first time. There was one new problem. Kraljic's strategy has eleven sections and fifteen smaller headings inside them, and on three runs out of four he rewrote the smaller headings in his own words. Describe a rule and he follows it sometimes. Write it out as a list and he follows it every time. So the list went into the file that is always in front of him, and four runs out of four kept every name.

Then came the part I am building the platform for. My agents share a task board. Every job is a card. We put a card on the board for Deepak and told him Kraljic's strategy was the card before it. The first thing Deepak did was open his own card, and the whole strategy was there, 29,000 characters. Four minutes later he had a preparation built on the highest price Kraljic said we could pay, and on what Kraljic said to do if there is no deal. Nobody copied and pasted anything.

Then Kahneman moved in a morning, Rackham in an afternoon, the Compiler in an evening. The whole buying side ran on the board, start to finish. So did the selling side. Then we asked Rackham, the seller's agent, to write the buyer's strategy. He refused in one line, and on his own he put a card on the board for Kraljic. Kraljic did it in three minutes.

All of this runs on a server at €12.99 a month and a ChatGPT subscription at €84.43.

You do not need a server for the two checks that did the work this week.

Ask your AI to prepare the same negotiation twice, in two fresh chats. What comes out the same both times is what it knows. Where it changes, the decision is yours. In our case the first offer changed every time.

And before you trust that it read the forty pages you pasted, ask it to quote one sentence from page thirty, without looking it up. If it cannot, it did not read them, whatever it told you.

Every negotiation becomes a fight over price. Not with AI

That is the first of four free sessions I run on Maven. It is the same $18.5 million renegotiation Deepak ran five times this week. We score how the buyer prepared against the five questions that decide a deal. She is good at her job. She gets two out of five. Then AI prepares the same deal again, both sides, in front of you.

Thursday 10 September, 18:30 Rome, 12:30 New York. 45 minutes with questions, nothing to prepare.

https://maven.com/p/02d6f7/every-deal-becomes-a-fight-over-price-not-yours

When you prepare with AI, have you ever asked it the same question twice and compared the two answers?

Just hit reply and tell me. I read and answer every one.

#Negotiation #NegotiationSkills #NegoAI #AI #AIAgents #AugmentedNegotiator