Dear Negotiation Explorer,
Welcome to issue 41 of our NegoAI series.
Last week, in Part 1, I built a working AI negotiation prep agent live in two different tools. One assistant. Two drawers — System Instructions and Knowledge Base. That issue ended on a question I left open: what happens when one assistant runs out of road?
Monday was the answer. Same case, but this time the assistant got replaced by a chain — four specialized agents in sequence — and the brief got handed to a counterparty simulator that played the partner across the table, live, for a stress-test. Part 1 was the foundation. Part 2 was the building.
And it lands one week before Cohort 2 of Build Your AI Negotiation System starts — Monday, May 25. That's the five-week version of what you're about to read, on your own deals. I'll come back to that at the end. First, here's what the chain produced and how it held up under pressure.
The Ceiling of One Assistant
In Part 1, the one-assistant version produced a quite-acceptable brief. Good enough for daily B2B work. But you could feel where it stopped — hidden interests, the counterpart's BATNA, multiple scenarios with diagnostic questions, integrative options that change the shape of the deal rather than the price.
Not a model failure. A single-instance limit: one agent juggling procurement strategy, counterpart psychology, scenario generation, and synthesis can't go deep in any of them. So you break the work apart.
The Chain — Four Specialized Roles
Same case as Part 1. Sarah Chen, Chief Procurement Officer at a pharma company. $18.5M consulting contract with McKinsey, a decade in. Senior partners scarce, juniors running the work, implementation below benchmark, competitors offering 25–38% less for similar scope. Sarah's objectives: a 25% cost reduction, a shift from hourly billing to outcome-based pricing, fewer but more senior consultants, and — non-negotiable — the relationship preserved. Not a price fight. A restructuring.
The chain has four roles:
Kraljic — procurement strategist. Classifies the supplier relationship using the Kraljic matrix and weights the issues. For this case it landed on outcome-based billing first, senior partner involvement second, price third, AI integration fourth — plus a stakeholder map and the trust-building moves to run alongside the commercial conversation.
Kahneman — counterpart psychology. Reads the partner's LinkedIn profile and produces a behavioral profile. David Morrison — head of US operations at McKinsey — came back ambitious and analytical, competing as the primary conflict mode at 70%. Expect a numbers-driven negotiator who respects substance and won't move on relational appeals alone.
Deepak — negotiation strategist. The synthesizer. Takes both upstream reads and generates three scenarios for what Morrison might be playing — collaborative, defensive, board-level bypass — each with a reservation value and the diagnostic questions that tell you which one is actually live. Then a value-creation pass (seven prompts run in sequence) and the power, leverage, and influence maps.
Compiler — the final agent. Not a text generator. A full agent with its own System Instructions and Knowledge Base, whose job is to stitch all three upstream outputs into the one-page brief you'd walk into the room with.
The architecture isn't linear. Kraljic and Kahneman run in parallel — Kahneman only needs the LinkedIn profile, so it doesn't have to wait for Kraljic. Both feed Deepak. All three feed the Compiler.
Case + LinkedIn profile
│
┌──────┴──────┐
▼ ▼
Kraljic Kahneman ◄── run in parallel
└──────┬──────┘
▼
Deepak ◄── synthesizes both
▼
Compiler ◄── reads all three
▼
Prep briefThat parallel-then-merge shape is what makes the chain fast and what makes the synthesis sharp. Each agent does its thing well, on its own ground, and Deepak does the work that requires both perspectives — the integration the single assistant couldn't carry.
The Pressure Test — Live, Against Von Neumann
The Compiler's brief is the plan. Whether the plan holds in the room is a different question. That's what Von Neumann is for — the counterparty simulator you read about a couple of issues back.
I uploaded two documents to a fresh Von Neumann chat: the original case, and the Compiler's brief. Von Neumann read both, picked the difficulty level (4 — appropriate for an executive-level negotiation), set the scene — PharmaCrest executive conference room, the room overlooks the New Jersey campus, David Morrison is already sitting at the table — and waited for Sarah to walk in.
I opened where the brief told me to: outcome-based pricing first, fewer but more senior consultants, AI displacing the analyst-tier work, relationship preserved.
Morrison didn't push back on the framing. He pushed back on the math.
He listened, took one note on his tablet, and said something close to: "I want to make sure I'm reconciling this. You're asking for a shift to our most premium senior-level resources, you're asking us to take on your implementation risk through an outcome-based model, and simultaneously you're asking for a 25% reduction in fee. Those three levers don't move in that direction at the same time."
That's the moment. In a real room, that's the moment the prep starts paying off — or stops. Because what Morrison just did is what a competent counterpart always does: he reframed three concessions as one impossible ask, and put the burden of proof on Sarah. Help me understand what's driving the 25%. Is this an internal mandate from finance, or are you benchmarking it against something specific?
The brief had the answer ready. Competitor offers in the 25–38% range; staying at the bottom of the bracket as a credit for McKinsey's institutional knowledge; implementation rate below benchmark; senior partners absent; AI absorbing the work the junior tier used to do.
Morrison's counter — and this is the part Deepak's value-creation pass had already anticipated — wasn't to fight on price. It was to reframe the deal. He proposed a co-development structure: trade upfront fees for shared AI innovation, align McKinsey's compensation with implementation success, restructure the $18.5M baseline into a model with significant upside performance bonus.
That's an integrative move. It's also exactly the kind of move Deepak's "future relationship + risk-sharing + cross-issue restructuring" prompts had already surfaced as a high-probability path. The plan held. Not because it scripted the conversation — it can't — but because it had already mapped the territory where the conversation could go.
What Repeats — and What This Means
Strip the demo back and there's one architectural fact under all of it: each of the four agents — Kraljic, Kahneman, Deepak, Compiler — has its own harness. Its own System Instructions. Its own Knowledge Base. Same separation we built in Part 1 for a single assistant, now applied four times.
That's the leverage. The thing you built in Part 1 isn't replaced by Part 2 — it's repeated. The six pillars of harness engineering (Context, Memory, System Instructions, Knowledge Base, Evaluation, Tools) don't change. They just stack per agent. Four agents in a chain is four times the same pattern.
And the pattern is tool-agnostic on purpose. I ran the workflow in Cassidy because Cassidy is a no-code visual builder and the chaining is easy to see on stage. But the same architecture runs in Microsoft Copilot — slower, because you hand each agent's output to the next manually — and it runs in a full Copilot Studio automation if your IT team is willing to wire it. The chain is the asset. The tool is a runtime.
The reason to do any of this — single assistant or chain — is what my research with 120 senior executives produced last year: 48% higher individual deal value, and 84% higher joint gains when both sides used it (work I presented at the Harvard Kennedy School AI Negotiation Forum in January 2026). The single-assistant version of the method already gets you there. The chain is where the work gets sharper.
What This Means for You
Don't try to build four agents on your first try. Build two.
The first agent is Deepak — the synthesizer. That's the spine; everything else feeds it. The second agent is your specialization — and it's a choice based on what your deals actually need:
Behavioral — if your counterpart's psychology is what you most need to read (sales, complex stakeholder negotiations, board interactions). This is the Kahneman role.
Specialist — buyer-side or seller-side. If you negotiate procurement contracts, this is a Kraljic-style agent. If you're a seller, it's a competitive-positioning agent tuned to your industry.
Simulator — if your pressure point is rehearsal, not preparation. Build a Von Neumann instead of a behavioral profiler, and use the chain as a sparring partner.
Whichever second agent you pick, both feed a text generator — a Compiler-lite — that turns the two perspectives into a one-page brief. That's a workable starting chain: two agents plus a generator. You don't need Cassidy. You don't need code. ChatGPT projects, Claude projects, Copilot agents — any of them will hold a two-agent chain. Same separation per agent: KB outside SI, four times. (Or, in the starter version, twice.)
That's what Cohort 2 of the course builds, end to end, on your own deals — with five weeks to refine the chain on the deal you're actually negotiating.
What's on the Calendar
Cohort 2 — Build Your AI Negotiation System, starts Monday, May 25. Five weeks, live, hands-on. You build your two-agent chain — Deepak plus your second agent (behavioral, specialist, or simulator) — feeding a text generator for the one-page brief you'd actually walk into a room with. No code. Any tool. The course is the build; Monday's session was the architecture peek. Alumni reviews on the landing page (Roar at Waegger Negotiation Institute; Yvonne at Coanea GmbH; and more).
Use code Nego25 at checkout for 25% off. This is the last issue before cohort 2 starts.
Details and registration: https://maven.com/nego-ai/build-your-ai-negotiation-system
Missed Part 2? Recording: https://maven.com/p/c70efd/build-your-ai-negotiation-workflow-no-code
Missed Part 1? Recording: https://maven.com/p/d82a4e/build-your-ai-negotiation-system-no-code-any-tool
This Week's Question
If you built a two-agent chain on your next real deal — Deepak plus one specialist — what would the second agent be: a behavioral profiler, a buyer/seller specialist, or a counterparty simulator?
Reply and tell me — I read every response.