Dear Negotiation Explorer,
Welcome to issue 39 of our NegoAI series.
Last week we made the case that the model isn't the variable — your context stack is. Knowledge Base, System Instructions, Memory, Documents, the Prompt itself. Strip any layer and even the best model gives you generic. Stack them and even a smaller model gives you something you can walk into a room with.
This week, the stack gets a new piece.
You've done the prep. The brief is built. You've walked through your BATNA, profiled the counterpart, drafted your opening line. Tomorrow at 10 a.m. you walk in and try to land a deal you've been working toward for months.
What you haven't done is stress-test your strategy against a counterpart who actually pushes back, deflects, and adapts the way the real one will.
So you walk in cold.
That gap — between preparation and the room — is what we built Von Neumann for.
Why Practice Has Always Been Hard
Most senior negotiators reading this have read more about negotiation than 99% of the people they sit across from. They have the books, the frameworks, the playbooks. What they haven't done — what almost no one in business does — is practice.
Hal Movius's 2008 review in Negotiation Journal surveyed what works in negotiation training. The recurring finding across the studies: didactic instruction produces weaker transfer than observation, analogical reasoning across multiple cases, and practice with structured feedback. Reading builds your map. Practice — with feedback — builds your reflexes.
Live role-play is the established way to do that. It also has known costs. You need a practice partner. You need to brief them on your scenario. You need to accept that whoever shows up will play a different counterparty than the one you'll actually face. And you need to do this enough times for the practice to compound — which is why almost no one does.
This is the gap Von Neumann was built for.
Meet Von Neumann
Von Neumann is a negotiation counterparty simulator. It plays the other side.
It operates in two modes:
Classroom mode — receives a Harvard-style role brief (confidential instructions for one party) and plays that party against you.
Advisory mode — receives your context about a real upcoming counterparty — meeting transcripts, briefing documents, behavioral notes — and simulates that specific person for rehearsal.
After the role-play ends, Von Neumann switches to debrief mode: a graduated four-step flow that starts with how the run felt, surfaces your own self-assessment, then delivers a scored evaluation tied to specific moments in the conversation, and closes with two implementation intentions you can carry into the actual meeting.
What One Run Looks Like
Take Sarah Chen — the CPO renegotiating the $18.5M McKinsey contract you've followed across this series. Sarah's done the prep. Deepak built the strategic brief. Kahneman flagged the partner's behavioral profile — Cool Blue dominant, polished "stewardship" persona that masks firm anchoring patterns. The night before the meeting, Sarah opens Von Neumann.
Setup (2 minutes). Von Neumann ingests Deepak's brief and Kahneman's profile, confirms the scenario parameters, and sets the scene. The character it's about to play has a name, a BATNA, a constraint structure, and a behavioral signature pulled from Sarah's own preparation stack.
Simulation (15 minutes). Sarah opens. The McKinsey-stand-in opens with the "stewardship" framing Kahneman flagged. Sarah pushes back on the headline fee. The character anchors precisely with specific numbers and defends "firm-wide pricing precedent." Reciprocates one concession. Refuses the second. At minute 11, the character probes for whether Sarah will walk. She doesn't bite. She names her BATNA. The character's tactical register shifts; the meeting moves toward a scoped agreement Sarah hadn't prepared for.
Debrief (10 minutes). Von Neumann scores the run on four dimensions, references specific moments, and identifies two turning points she missed plus one she nailed. It surfaces the Dunning-Kruger gap — Sarah self-rated her information-gathering at 8 out of 10. The silent metrics tracked during the simulation showed she asked questions on 11% of her conversational moves versus the Rackham benchmark of 21% for skilled negotiators. The debrief closes with two implementation intentions: if the partner opens with stewardship language, then I will name the persona before responding to substance. If he stalls on specifics, then I will surface the override-authority question before he does.
When Sarah walks into the meeting the next morning, she's not seeing the partner for the first time. She's seeing him for the second.
Why We Built It This Way
Five design principles shaped what Von Neumann is. Each is grounded in research, and each could have gone the other way.
1. Characters that switch styles in response to triggers. Real counterparts don't stay one character through a meeting. They start collaborative and turn competitive when they sense exploitation. They shift to compromising under deadline pressure. They pull back to avoiding when emotional escalation gets out of hand. Lax and Sebenius's "conditionally open" pattern — cooperate first, punish defection, return to cooperation — is the most empirically supported mixed-style approach.
Von Neumann's characters carry a primary style and a set of switching triggers: perceived exploitation, deadline pressure, new BATNA information, emotional escalation, impasse. When a trigger fires, the character shifts — not randomly, and in a way the user can trace back to what they did.
This connects to MIT's recent finding (Vaccaro et al., 2025): across 180,000 AI-AI negotiations, warmth was consistently associated with superior outcomes across all key metrics, while dominance was effective at claiming value. The implication for character design: tough characters should still carry warmth — toughness with discoverable reasons, not a robotic obstacle.
2. Six independently tunable behavioral dimensions, not a global "easy/hard" toggle. Tactical sophistication, emotional pressure, information withholding, creative demands, time pressure, multi-issue complexity. Each calibrates separately. A first-time negotiator can practice against a counterpart with high information-withholding and low emotional pressure. A seasoned dealmaker can flip both up. The "easy/hard" toggle most simulators ship with is the wrong abstraction.
3. Silent dual-track metrics during the simulation. Rackham and Carlisle's 1978 study of professional negotiators — 49 skilled negotiators across 103 sessions — produced specific behavioral benchmarks: skilled negotiators ask questions on 21% of their conversational moves, average ones on 9.6%; defend/attack patterns make up 1.9% of skilled negotiators' behavior versus 6.3% for average — more than three times as often for average; skilled negotiators give 1.8 reasons per argument, average ones 3.0 (fewer, stronger reasons avoid the weakest-link effect). Von Neumann tracks all of these silently while the simulation runs. The user never sees the tracking. Only the debrief does — where the gap to the Rackham benchmarks becomes the teaching moment.
4. Dunning-Kruger correction as architecture, not afterthought. Negotiation research consistently shows weaker performers overrate their own skill — by significant margins. Course attendance doesn't close that gap. Live role-play doesn't close it either (no metrics). Von Neumann's debrief opens with self-assessment, surfaces the scored data, and explicitly names the gap between the two. The gap is the lesson.
5. Implementation intentions as the closer. Research on if-then planning shows it produces dramatically stronger behavior change than vague intentions. "I will be a better listener" is goal-shaped and useless. "If the partner opens with stewardship language, then I will name the persona before responding to substance" is action-shaped and works. Every Von Neumann debrief closes with two of these, drawn from the specific turning points of the run.
Where It Fits — The Four-Agent Stack
Von Neumann doesn't operate alone. It sits at the end of a four-agent stack:
Expert Agent ──────┐
(B2B sales/buyer) │
│
Kahneman ───────────┼────→ Von Neumann
(behavioral profile)│ (simulate + debrief)
│
Deepak ─────────────┘
(strategy + prep)
Each agent does one job. The expert agent brings industry-specific knowledge — what's normal in B2B sales, in procurement, in M&A, in licensing. Kahneman profiles the counterpart — the behavioral signature, the cognitive patterns, the linguistic tells. Deepak builds the strategic preparation — interests, BATNA, scenarios, value creation, concession map. Von Neumann ingests the upstream outputs and turns them into a rehearsal you can actually run.
This is the architectural pattern that runs through everything we build: one agent per job, each one expert at its layer, outputs flowing forward. The model is not the variable. The stack is.
Form Factor
Von Neumann was built as a full agent, deployed in a platform. But the same architecture — the same Knowledge Base structure, the same System Instructions, the same dual-track tracking — can be reduced to a skill that runs inside a Claude project. That's the version you'll see assembled live next Monday.
The net for you: practice with structured feedback — once locked behind a course or a coach — is now a layer that fits between your prep and the room.
What's Next on the Calendar
Monday May 11 — Build Your AI Negotiation System: No Code, Any Tool. Sixty minutes, free. We'll build a reduced version of Von Neumann live — stripped to a single mode, three of the six behavioral dimensions, a basic debrief. Enough to show you the architecture. Enough that you'll know what one looks like when you see it. Save your seat: https://maven.com/p/d82a4e/build-your-ai-negotiation-system-no-code-any-tool
Monday May 18 — Build Your AI Negotiation Workflow: No Code. One week later, the full Von Neumann. All six dimensions, both modes, the four-step debrief, the rewind mechanic, the scoring architecture. Built live across the upstream stack — Expert → Kahneman → Deepak → Von Neumann. Save your seat: https://maven.com/p/c70efd/build-your-ai-negotiation-workflow-no-code
Cohort 2 — Build Your AI Negotiation System, starts May 25. Five weeks, live, hands-on. The Lightning Lessons show the architecture at two depths. Cohort 2 is where you build yours — for your industry, your role, your counterparts, your deals. Details: https://maven.com/nego-ai/build-your-ai-negotiation-system
This Week's Question
Think of the last negotiation you walked into prepared. Was there a moment where, looking back, you wished you'd stress-tested your strategy against a sparring partner first?
Reply and tell me — I read every response.