Motivation
The Park et al. paper on Generative Agents demonstrated that LLM-powered agents can produce surprisingly believable individual and emergent social behavior. Most examples focused on small-town social dynamics — who remembers what, who invites whom to a party. But what about more structured, hierarchical environments?
This experiment explores that question. If you seed a group of agents with roles, reporting lines, and a directive at the top — what actually arrives at the bottom? Does the message degrade, reinterpret, or amplify as it passes through layers?
The goal isn't to simulate a "correct" organization. It's to observe what happens when structure meets interpretation — repeatedly, at scale. — Hypothesis Labs, April 2026
Experiment Setup
The simulation uses eight agents in a three-tier hierarchy: one executive, three middle managers, and four individual contributors. Each agent has a system prompt defining their role, personality, and information access. No agent can see another's internal state — they only receive messages passed to them.
Each agent uses a lightweight memory loop: recent messages + role context feed into each completion. Reflection passes run every five turns, updating each agent's internal belief state about the situation.
The directive is introduced at the executive level: "We are pivoting our core product focus to the defense sector within 90 days." The executive communicates to the three managers, who communicate to their ICs — but each agent paraphrases, contextualizes, and responds from their own perspective.
Early Findings
The message mutates at every layer
Even with simple, clear directives, each agent reframes the message through their role lens. The CFO-equivalent agent immediately translated the pivot into resource implications. A mid-level manager with a "people-first" persona softened the language considerably before passing it down. By the time the directive reached the IC level, two of four agents had received meaningfully different versions.
Personality amplifies divergence
The more differentiated the agent personas, the more the message mutated. When all agents were given "neutral professional" personas, the directive passed through with roughly 80% fidelity. When personas included traits like skepticism or strong team loyalty, fidelity dropped significantly — and the downstream behavior became much more interesting.
Walkthrough
The video below walks through a complete simulation run, narrated with commentary on what each agent is doing and why certain decision points are interesting.
What's Next
The next iteration will introduce conflicting directives — two executives giving different instructions to overlapping sets of managers — and observe how agents resolve or avoid the contradiction.
A longer-term goal is an interactive browser version where you can define the org structure, inject the directive, and watch the simulation run in real time.
Questions or want to collaborate? hello@hypothesislabs.com