Autonomous agents
Each employee maintains independent state and chooses actions locally, allowing team-level behavior to emerge without a central script.
Truuts experiment
The Office Experiment: Autonomous Agent Simulation
A live, browser-based office populated by autonomous pixel-art agents whose work, rest, social behavior and movement emerge from changing internal state and a shared 90-day deadline.
Tamashii simulation
Quick facts
- Creator
- Eugene Truuts
- Project type
- Autonomous agent simulation
- Core model
- Weighted utility decisions and A* navigation
- Simulation scale
- 1–24 agents across a 90-day project
- Runtime
- Browser · React · HTML Canvas
- Status
- Active experiment · Updated August 2026
About the experiment
The Office Experiment is an interactive agent-based simulation of a software team working toward a shared quarterly goal. It focuses on the small individual decisions that create larger organizational patterns: sustained focus, crowded coffee breaks, spontaneous conversations, exhaustion and deadline-driven crunch.
The live model starts with 12 agents, and the population control can rebuild the office with anywhere from 1 to 24. Developers, QA employees and managers move through the same physical workspace but receive different behavioral weights. Each character also has individual workaholic, social, anxious and resilience traits.
This is not a scripted animation and the agents are not driven by a language model. Tamashii is a transparent behavioral system: it repeatedly evaluates observable state, assigns weights to available activities and samples the next action. Small differences in personality, timing and position allow each run to develop differently.
The simulated project lasts 90 days. Work increases progress, continuous focus consumes energy and raises burnout, and falling behind increases urgency. At high urgency the office enters crunch mode: work becomes more likely and more productive, but recovery is suppressed and burnout accumulates faster.
How it works
Every agent moves through a continuous perceive–decide–navigate–act loop. Tamashii combines the shared project context with personal state, calculates a weight for each available activity and makes a probabilistic selection rather than following a fixed schedule.
The decision model considers:
- Deadline context: project progress, days remaining, urgency and crunch mode.
- Physical condition: current energy, accumulated burnout, focus duration and time since the last break.
- Role: developers, QA and managers have different baseline work, social and recovery tendencies.
- Personality: workaholic, social, anxious and resilience values modify those role-level tendencies.
- Available activities: focused work, coffee, eating, conversation, wandering and celebration after project completion.
- Space: selected intentions become destinations, then A* pathfinding routes characters around walls and furniture on the office navigation grid.
Debug information is enabled by default so the simulation remains inspectable. Labels expose current activity, goal, energy and burnout; colored paths show planned routes, destinations and blocked navigation cells.
Play the experiment
A standalone playable version of The Office Experiment is available on itch.io.
Technical concepts
Weighted utility decisions
Tamashii scores work, coffee, food, conversation and recovery, then samples from those weights so identical conditions do not always produce identical behavior.
Energy and burnout
Focused work drains energy and accumulates burnout. Breaks restore capacity, while resilient personalities tolerate pressure for longer.
Deadline pressure
A 90-day project clock compares remaining work with remaining time. Rising urgency changes priorities and can activate a high-output, high-cost crunch state.
Roles and personality
Developer, QA and manager baselines are modified by individual workaholic, social, anxious and resilience values.
A* pathfinding
Intentions become physical destinations. Agents calculate collision-aware routes across a 55 × 44 navigation grid and move around walls and furniture.
Emergent office behavior
Agents share the same environment, so independent choices produce visible rhythms: work clusters, social encounters, queues for resources and collective celebration.
Inspectable simulation
The debug overlay exposes observable inputs, goals and routes. It explains system behavior without pretending to reveal hidden chain-of-thought reasoning.
Simulation insight
Watch the feedback loop between productivity and recovery. More agents can accelerate progress, but the result is not a simple head-count equation: roles, travel time, depleted energy and burnout all change how much of the population is actually working.
Agent stateActivity · goal · energy · burnout
Project contextDay · progress · urgency · crunch mode
Navigation
Colored routes connect each moving agent to a selected destination; red cells mark non-walkable space.
Decision model
Role and personality modify weighted choices, while visible thought and speech bubbles reflect current activities.