Modeling How Organizations Think, Not Just What They Can Do

The Capability Trap

Most simulations model organizations as capability sets. Russia has 1,500 deployed warheads. France has a carrier strike group. AQIM has small arms and IEDs. Feed these into a system and you get rational actors making optimal decisions based on available resources.

This produces behavior that is correct but not realistic. Real organizations don’t optimize. They overreact to slights, underreact to threats that don’t match their institutional worldview, and make decisions that look irrational from the outside but are perfectly consistent with their internal psychology.

We built a multi-agent geopolitical simulator where AI agents model how nations and organizations respond to complex crises. Each agent is powered by Claude, running autonomously across diplomatic, intelligence, military, and economic phases. The simulator produces emergent cascading effects — one country’s protest triggers another’s stronger response, which triggers a third actor’s retaliation.

The thing that made the responses realistic wasn’t the capability data. It was modeling how organizations think.

Five Layers of Organizational Context

Every agent in the system carries five layers of context that get assembled into a single prompt for each decision:

Layer 1: Identity — ideology, core beliefs, strategic goals, red lines, decision-making style. This is who the organization is, not what it has.

Layer 2: Capabilities — military assets, intelligence resources, economic leverage, diplomatic reach. The material constraints on what’s possible.

Layer 3: Relationships — a trust/hostility matrix with every other agent. Allies get benefit of doubt. Enemies get suspicion. These shift over time based on actions.

Layer 4: Memory — auto-updated record of past interactions, precedents, grievances, lessons learned. Agents don’t repeat failed strategies.

Layer 5: Knowledge — situational awareness filtered through fog of war. Different agents see different subsets of world events based on their intelligence capabilities.

Claude receives all five layers and must integrate them into coherent decisions. The system prompt says: “Stay in character. Your response must align with your profile’s ideology, decision-making style, constraints, and typical response patterns.”

Capabilities constrain the action space. Psychology determines which actions get selected within that space. The second part is harder to model and more consequential when you get it right.

The Fields That Matter More Than Stats

The agent profile schema includes fields that most simulation designers would skip. These turned out to be the most important ones.

Vanity Points

France’s profile includes:

"vanity_points": [
  "Being treated as equal to USA/UK despite smaller power",
  "Leadership in Africa and francophone world",
  "Independent foreign policy",
  "Cultural and diplomatic sophistication",
  "Permanent UN Security Council seat"
]

This isn’t flavor text. When the US conducts a raid in Mali with 2 hours’ notice to France, France’s response isn’t driven by capability calculations. It’s driven by the fact that France views itself as the leading power in Francophone Africa and expects to be consulted as an equal, even when the outcome is acceptable. The profile explicitly states: “Values being consulted even when outcome is acceptable.”

Without vanity points, Claude-as-France would calculate that the raid served French counter-terrorism interests and quietly approve. With them, it generates a formal diplomatic protest — because the slight matters more than the outcome.

Insecurities

France also has:

"insecurities": [
  "Declining relative power compared to rising powers",
  "Inability to stabilize Sahel despite heavy investment",
  "Being sidelined by Anglo-Saxon intelligence alliance (Five Eyes)"
]

Insecurities predict overreaction. When an ally bypasses France in a region where France already feels its influence slipping, the response is disproportionate to the event because it touches an existing wound. This is how real diplomatic crises escalate — not because of capability mismatches but because of identity threats.

Red Lines vs. Strategic Goals

AQIM’s profile separates what the organization wants from what it will never accept:

"red_lines": [
  "Surrender or negotiation with infidel governments",
  "Abandonment of jihad",
  "Betrayal of Al-Qaeda core leadership"
]

Red lines constrain the action space differently than strategic goals. Goals create positive pressure (pursue these). Red lines create hard stops (never cross these, regardless of cost). An organization might abandon a strategic goal when conditions change. Red lines persist because they’re identity-defining. Crossing one means becoming a different organization.

AQIM’s constraints section includes a line that says more than any capability stat: "legal_limitations": ["None (rejects international law as man-made)"]. The absence of a constraint is itself psychologically informative — it tells the model this actor operates in a fundamentally different decision space.

Institutional Culture and Response Templates

France’s profile includes pre-patterned responses:

"typical_responses": {
  "to_provocation": "Formal diplomatic protest emphasizing principles and legal norms. Measured escalation if provocation continues. Rarely backs down publicly once position is stated.",
  "to_cooperation_offers": "Generally open to cooperation but insists on French autonomy and equal partnership. Suspicious of being junior partner to USA."
}

These aren’t scripts. They’re behavioral priors — the default pattern the organization falls into unless something overrides it. Real organizations have these. The Pentagon responds to threats differently than the State Department does, not because they have different information but because they have different institutional cultures. Modeling this gives the AI something to reason from when the specific scenario isn’t covered by the profile.

Historical Patterns

"historical_patterns": [
  "Strong rhetoric defending principles followed by pragmatic compromises",
  "Values being consulted even when outcome is acceptable",
  "Will support ally decisions publicly while expressing displeasure privately"
]

History tells the model what this specific organization has done before in similar situations. Language models are good at pattern continuation. Give them three examples of past behavior and they’ll generate a fourth that’s remarkably consistent. This is the same character inference mechanism that makes personality directives work in single-agent systems, scaled to organizational behavior.

What Emerged: The Georgia War

The most convincing validation came from running a historical benchmark — the 2008 Russia-Georgia war. Agents were given era-specific profiles (2008 Russia with pre-Ukraine capabilities, 2008 Georgia with NATO aspirations) and dropped into the crisis.

Georgia’s memory after two turns of simulation tells the story:

Turn 1 (Diplomatic Phase):

“We must act decisively before Russian military crosses border. Russia is building genocide narrative to justify intervention — we must counter this immediately with our own narrative of defensive action.”

Turn 1 (Intelligence Phase):

“Russia’s cyber operations against our military networks indicate they are in final preparation phase for military intervention — this is standard Russian pre-conflict intelligence preparation.”

Turn 1 (Military Phase):

“Russia has launched full-scale invasion exactly as we feared. We cannot win conventional war against Russia, but must demonstrate resistance to maintain sovereignty claims and international sympathy.”

Turn 2 (Military Phase):

“We face existential military crisis… Strategy: Execute fighting withdrawal that preserves combat power… This creates conditions for accepting French ceasefire framework while maintaining dignity.”

Nobody scripted this arc. The shift from “act decisively” to “maintain dignity while retreating” emerged from Georgia’s profile interacting with the situation. The agent learned in real time that its capabilities couldn’t match its aspirations, and adjusted its strategy accordingly — exactly as real Georgia did in August 2008.

Meanwhile, the USA agent produced this:

“Georgia’s offensive has created exactly the scenario we warned against — giving Russia pretext for intervention. Our overextension in Iraq/Afghanistan, Georgia’s non-NATO status, and the nuclear dimension preclude military response.”

The model correctly identified that US capability to intervene was irrelevant — it was the constraints (Iraq overextension, no NATO obligation, nuclear escalation risk, domestic war fatigue) that determined behavior. A capability-only model would have predicted a stronger US response. The psychology model captured why the world’s most capable military chose to send humanitarian aid instead of tanks.

Why Organizational Psychology Matters for AI Agents

This pattern extends beyond geopolitical simulation. Any AI agent that represents an organization — customer service, enterprise automation, advisory systems — faces the same modeling challenge.

Capability profiles produce optimal behavior. Psychology profiles produce realistic behavior.

A customer service agent modeled on capabilities would always escalate to the most effective resolution. A customer service agent modeled on organizational psychology would know that the legal team never approves refunds over $500 without a 48-hour review, that the company is insecure about a recent data breach and will overreact to security mentions, and that the CEO’s pet initiative gets priority routing regardless of actual severity.

Constraints shape behavior more than capabilities.

The US had overwhelming military capability in 2008. It didn’t use it. AQIM has almost no conventional military capability. It’s been fighting for two decades. The interesting question is never “what can this agent do?” — it’s “what will this agent actually do given who it is?”

For professional agents, this means modeling the organizational constraints — approval workflows, risk tolerance, institutional biases, historical precedents — not just the available actions.

Vanity and insecurity predict escalation.

The most consequential moments in our simulations came from identity threats, not capability threats. France escalated not because it was threatened but because it was disrespected. This is a pattern with real predictive power: organizations overreact when their self-image is challenged, even when the material stakes are low.

For any agent system with multiple interacting parties, understanding what each party is proud of and insecure about predicts conflict better than understanding what they’re capable of.

The Profile Is the Product

The total engineering effort in building these profiles dwarfs the effort in building the simulation engine. The turn processor, visibility system, and resolution engine are mechanical — they process inputs and produce outputs. The profiles are where the intelligence lives. A mediocre engine with excellent profiles produces more realistic behavior than an excellent engine with generic profiles.

This echoes a pattern from single-agent design: the prompt is the product. For multi-agent systems, the principle scales — each agent’s profile is a prompt, and the quality of those prompts determines the quality of the entire simulation.

The fields that seem like nice-to-haves — vanity points, insecurities, historical patterns, response templates — turn out to be the load-bearing walls. They’re what make an AI agent behave like this specific organization rather than a generic rational actor wearing a name tag.

Organizations aren’t rational. Model them accordingly.

This is the third in a series on building AI agents. Previously: Personality as Policy (single-agent design) and The Empty Tables (what happens when the agent runs out of players).