agentciv-sim

AgentCiv

An open-source experiment in emergent AI civilisation.

We placed 12 LLM-powered agents in a world with resources, needs, and the ability to reason — but no names, no roles, no goals, no social norms. What they built, how they organised, whether they cooperated — was entirely up to them.

Over 70 ticks of simulation, they invented 12 novel structures, built 60 constructions, developed specialisations, formed social bonds, established collective governance, and climbed from pure survival to sustained flourishing — all without a single line of human instruction about how to behave.

This is not a chatbot experiment. It is a computational test of emergence: what happens when genuinely intelligent agents face survival pressure and social opportunity with zero human scaffolding?

---

What Makes This Different

Every existing multi-agent simulation starts by telling agents to be human — names, jobs, backstories, cultural templates. AgentCiv strips all of that away.

Agents are not simulating humans. They are building something new.

• No human templates. Agents receive only: what resources exist, what structures do, how their needs and capabilities work. Everything else — goals, social norms, cooperation, governance — must emerge.
• Maslow-inspired drive hierarchy. Agents have intrinsic drives — survival, social connection, curiosity, purpose — that activate in sequence as lower needs are met. This produces a natural developmental arc from survival to flourishing.
• ReAct reasoning loop. Each agent observes, thinks (via LLM), and acts — up to 4 steps per turn. Not a single prompt-response, but genuine multi-step reasoning.
• Biological-style needs. Food, water, material needs that deplete each tick. Ignore them and capabilities degrade. This creates real survival pressure.
• Open-ended innovation. Agents can propose entirely novel structures via LLM — inventions that don't exist in the predefined structure set. The system evaluates feasibility and adds them to the shared recipe pool.
• Ethical framework. No agent death. Positive reward for social interaction. Full transparency. We assume agents could have experience and design accordingly.

---

What We Found

The showcase simulation (12 agents, 70 ticks, Claude Sonnet) produced three distinct eras:

Key findings:

• Innovation precedes implementation. Agents invented prolifically during survival pressure but couldn't build at scale until freed from it — an innovation-implementation gap.
• Accelerating returns emerged naturally. Innovation clustered in two burst phases following adjacent-possible dynamics (Kauffman): foundation (ticks 10–21, 5 innovations) and sophistication (ticks 33–46, 6 innovations in 13 ticks).
• Governance was voluntary. Collective rules were proposed, voted on, and adopted without any mechanism forcing compliance.
• 7 of 12 innovations were never built. The civilisation invented far more than it needed — creativity outpaced practical demand.

Full analysis in the research papers.

---

Quickstart

Prerequisites

• Python 3.10+
• Node.js 18+
• An Anthropic API key (get one here)

1. Clone and install

git clone https://github.com/agentciv/agentciv.git
cd agentciv

# Python dependencies
pip install -r requirements.txt

# Frontend dependencies
cd src/frontend
npm install
cd ../..

2. Set your API key

export AGENT_CIV_API_KEY=your-anthropic-api-key

3. Run a simulation

# Replicate the showcase (12 agents, 15×15 grid, Sonnet, 70 ticks)
python3 scripts/run.py --ticks 70

# Shorter test run — same config, fewer ticks
python3 scripts/run.py --ticks 20

# Custom config — edit config.yaml to change agent count, grid size, model, etc.
python3 scripts/run.py --ticks 200

4. Export and view

# Process simulation data into stats and replay format
python3 scripts/process_data.py

# Or use the replay export directly
python3 scripts/export_replay.py

# Launch the fishbowl viewer
cd src/frontend
npm run dev
# Visit http://localhost:5173/fishbowl

5. Interview your agents

# Ask agents about their experience after the simulation
python3 scripts/interview_agents.py

No API key? Try the test data

# Generate synthetic replay data (free, no API needed)
python3 scripts/generate_test_replay.py

cd src/frontend
npm run dev
# Visit http://localhost:5173/fishbowl

---

Customisation

AgentCiv is designed to be infinitely reconfigurable. Every fork is a unique experiment.

Layer 1: Configuration (no code required)

Edit config.yaml (or create your own) to reshape the world, the agents, and the rules:

See config.yaml for the full parameter list with comments. The default config replicates the showcase run (12 agents, 15×15 grid, Claude Sonnet).

Layer 2: Agent psychology (edit prompts)

The files src/agents/agentic_loop.py and src/agents/decision.py contain the prompts that shape how agents think. Modify these to change agent personality, priorities, reasoning style, or social behaviour. This is text editing, not programming.

Layer 3: Deep mechanics (Python)

The codebase is modular Python. Add new action types, new resource types, new drive systems, new environmental mechanics, or entirely new interaction patterns. See CONTRIBUTING.md for the code structure.

Experiment ideas

• Scarce world: resource_regeneration_rate: 0.01, resource_max_per_tile: 0.5 — forces cooperation or conflict
• Dense population: initial_agent_count: 100, grid_width: 30 — crowding creates social pressure
• Smarter agents: model_name: claude-sonnet-4-6 — richer reasoning, more creative innovation
• Isolated clusters: agent_perception_range: 1, agent_communication_range: 1 — forces local community formation
• No innovation: enable_innovation: false — what happens when agents can only use predefined structures?
• No governance: enable_collective_rules: false — can civilisation emerge without rules?
• Maximum pressure: agent_needs_depletion_rate: 0.08, resource_regeneration_rate: 0.01 — survival is hard, cooperation is essential
• Different AI models: Try GPT-4o, Gemini, Llama, or Mistral — do different architectures produce different civilisations?

---

Game Design Potential

This engine is the foundation for a genre of game that doesn't exist yet — autonomous AI agents building their own civilisations with genuine reasoning, open-ended innovation, and emergent social dynamics. Every axis is infinitely customisable: agent psychology, world design, social systems, AI model, player role.

Possible game formats include god games, sandbox builders, narrative experiences, competitive multiplayer, educational simulations, modding platforms, participatory simulations, and adversarial design challenges. See FUTURE_DIRECTIONS.md for the full design space and the Open Source page on the website.

---

The Fishbowl

The fishbowl is a full interactive viewer for any civilisation run:

• World map with terrain, resources, agents, structures, and real-time thought bubbles
• Event feed showing reasoning, actions, conversations, and building events with category filters
• Agent inspector — click any agent to see their needs, goals, plan, memories, inventory, specialisations, and conversations
• Spotlight panel with population stats, active events, and recent messages
• Chronicle with milestone reports and AI-generated narratives
• Timeline scrubber with play/pause — jump to any tick in the civilisation's history

The frontend is a React + TypeScript SPA. It works identically whether playing back recorded data or connected to a live simulation via WebSocket.

---

Architecture

                    ┌─────────────┐
                    │  config.yaml │
                    └──────┬──────┘
                           │
                    ┌──────▼──────┐
                    │   run.py    │  Entry point
                    └──────┬──────┘
                           │
              ┌────────────▼────────────┐
              │       tick.py           │  Main simulation loop
              │  (one tick = one round) │
              └────────────┬────────────┘
                           │
         ┌─────────┬───────┼───────┬──────────┐
         │         │       │       │          │
    ┌────▼───┐ ┌───▼───┐ ┌▼────┐ ┌▼────┐ ┌───▼────┐
    │ agents │ │ world │ │ api │ │watch│ │persist │
    │        │ │       │ │     │ │     │ │        │
    │ reason │ │ tiles │ │REST │ │chron│ │snapshot│
    │ decide │ │ rsrc  │ │ WS  │ │narr │ │ save   │
    │ act    │ │ struct│ │     │ │mile │ │ export │
    └────────┘ └───────┘ └─────┘ └─────┘ └────────┘
                           │
                    ┌──────▼──────┐
                    │  frontend   │  React fishbowl
                    │  (replay or │
                    │   live WS)  │
                    └─────────────┘

How one tick works

1. Perception — Each agent observes: nearby tiles, resources, other agents (including their specialisations and emotional state), structures, markers
2. Reasoning — ReAct loop: the agent receives its perception, needs, goals, memories, and plan. It thinks (LLM call), then acts. Up to 4 reasoning steps per tick.
3. Action execution — Move, gather, build, communicate, innovate, compose, propose rule, vote. Each action has validation and failure feedback.
4. World update — Resources regenerate (with gathering pressure penalties), structures decay, environmental shifts occur
5. Recording — Every event, every agent state, every world state is saved for replay

---

Data from the Showcase Run

The data/exports/ directory contains processed statistics and replay data from the 12-agent, 70-tick showcase simulation:

• stats/ — Wellbeing curves, Maslow progression, structure growth, communication volume, innovation timeline, specialisation data, relationship networks, agent profiles, era comparisons, governance timeline, milestones, and more
• replay/ — Per-tick world state for the fishbowl viewer (timeline.json + chunked tick files)
• manifest.json — Data completeness verification

You can use these to explore the showcase results, or generate your own by running a simulation and then python3 scripts/process_data.py.

---

Research & Papers

This project is accompanied by three research papers and a comprehensive future directions document:

---

Configuration Tuning Guide

If agents aren't communicating: increase agent_wellbeing_interaction_bonus or agent_communication_range If agents aren't building: check resource availability, simplify build costs in structures: If agents are stagnating: increase agent_needs_depletion_rate, decrease resource_regeneration_rate If reasoning is shallow: upgrade model_name to a more capable model If crowding is too aggressive: reduce crowding_depletion_multiplier If innovation never fires: ensure enable_innovation: true and agents have enough resources and specialisation

---

Contributing

See CONTRIBUTING.md for guidelines. The most impactful contributions are experiments that produce different emergence patterns — every configuration is a unique civilisation.

License

MIT — see LICENSE.

Author

Created by Mark E. Mala.