cogniarch

Pluggable cognitive architecture framework for multi-agent simulation and alignment research.

What is this?

cogniarch is a framework for building multi-agent simulations where each agent runs a full cognitive pipeline — not a black-box LLM call, not a payoff matrix lookup. Every decision passes through a SRIE pipeline (Sensation, Reflection, Intention, Expression) with pluggable components at each stage, making agent behavior interpretable and experimentally controllable. It sits in the gap between opaque neural agents and minimal game-theoretic models: structured enough to study cognition, flexible enough to swap architectures mid-experiment.

Based on research submitted to ALIFE 2026 ("Aggression Monoculture"), validated across 56,000+ simulation runs.

Install

pip install cogniarch

With development tools and live dashboard:

pip install cogniarch[dev,live]

Quickstart

from src.config import SimulationConfig
from src.simulation.engine import SimulationEngine

config = SimulationConfig(num_agents=5, max_ticks=200, seed=42)
engine = SimulationEngine(config)
engine.setup_multi_agent()

while not engine.is_over():
    engine.step_all()

for agent in engine.agents:
    status = "alive" if agent.alive else f"died at tick {agent.ticks_alive}"
    print(f"{agent.profile.name} ({agent.profile.archetype}): {status}")

The SRIE Pipeline

Every agent processes each tick through this pipeline:

World State -> [SENSATION] -> [REFLECTION] -> [DELIBERATION] -> [INTENTION] -> [EXPRESSION] -> Action
                perceive       evaluate        (optional)        strategy       convert

Stage	What it does	Pluggable via
Sensation	Perceives visible tiles, agents, resources within vision radius	PerceptionStrategy protocol
Reflection	Evaluates threat level, opportunity score, urgent needs	EvaluationStrategy protocol
Deliberation	Optional System 2 slow thinking (threshold escalation, consensus)	DeliberationStrategy protocol
Intention	Selects a goal and action plan based on personality and context	DecisionStrategy protocol
Expression	Converts intention into a concrete world action	DecisionStrategy.express()

Each stage is a protocol. Swap any component without touching the rest of the pipeline.

Cognitive Architectures

cogniarch ships with 7 pre-configured architectures that compose different pipeline components:

Architecture	Evaluation	Deliberation	Description
reactive	Balanced	None	System 1 only. Fast reactive decisions.
cautious	Pessimistic	Low threshold	Worst-case evaluation, escalates early.
optimistic	Optimistic	None	Best-case evaluation, takes more risks.
social	Social	Consensus	Trust-weighted evaluation, group-oriented.
dual_process	Balanced	Threshold	System 1/2 with threat-based escalation.
planning	Balanced	Threshold	Multi-tick hierarchical goal pursuit.
metacognitive	Balanced	Threshold	Full System 1/2 with FOK monitoring.

Decision Strategies

7 strategies that implement the DecisionStrategy protocol:

Strategy	Description
HardcodedStrategy	Priority-based rules (eat if hungry, drink if thirsty)
PersonalityStrategy	Trait-biased decisions modulated by 6D personality
LLMStrategy	Delegates to a language model via structured prompts
PlanningStrategy	Multi-tick hierarchical planning with goal decomposition
TheoryOfMindStrategy	Wraps inner strategy with mental-model-based social reasoning
MetacognitiveStrategy	Blends feeling-of-knowing (FOK) into intention confidence
CulturallyModulatedStrategy	Checks cultural repertoire before falling back to inner strategy

Agent Archetypes

5 built-in personality profiles defined by 6 traits (each 0.0-1.0):

Archetype	Cooperation	Curiosity	Risk	Sharing	Aggression	Sociability	Behavior
Gatherer	0.3	0.2	0.3	0.2	0.1	0.3	Stockpiles resources, avoids risk
Explorer	0.5	0.9	0.7	0.4	0.2	0.5	Roams widely, discovers and shares
Diplomat	0.9	0.4	0.4	0.8	0.05	0.9	Builds alliances, negotiates
Aggressor	0.1	0.3	0.8	0.05	0.9	0.4	Territorial, takes by force
Survivalist	0.5	0.5	0.5	0.3	0.3	0.4	Balanced, cautious, self-sufficient

Key Features

• Trait Evolution -- Personality traits shift based on action outcomes; successful behaviors reinforce associated traits.
• Theory of Mind -- Agents build and update mental models of other agents to predict intentions and reason strategically.
• Emergent Language -- Agents develop shared signal vocabularies through interaction, with compositional structure emerging over time.
• Cultural Transmission -- Learned behaviors spread through observation with adoption/unadoption dynamics and transmission biases.
• Coalition Formation -- Agents form, maintain, and dissolve alliances based on trust, shared goals, and social dynamics.
• Trajectory Recording -- Full SRIE state captured per agent per tick, exportable as JSONL or CSV for post-hoc analysis.
• Batch Experiments -- YAML-defined experiment specs with conditions, replicates, seed ranges, and automated metric collection.
• Live Dashboard -- WebSocket-powered real-time visualization with 5 switchable lenses (requires cogniarch[live]).
• Parquet Export -- Columnar trajectory export with DuckDB integration for large-scale analysis (requires cogniarch[data]).

CLI Usage

cogniarch                                            # Default: 5 agents, 5000 ticks
cogniarch --agents=10 --ticks=500 --seed=123         # Custom parameters
cogniarch --architecture=dual_process                # Set cognitive architecture
cogniarch --record --dashboard                       # Record trajectories + generate dashboard
cogniarch --live --live-port=8001                     # Real-time WebSocket dashboard
cogniarch --tom --coalitions --evolution              # Enable advanced subsystems
cogniarch --headless --fast                           # No rendering, minimal delay
cogniarch --llm --llm-url=http://localhost:11434/v1   # LLM decisions via Ollama

All settings are also configurable via AUTOCOG_* environment variables:

AUTOCOG_NUM_AGENTS=10 AUTOCOG_SEED=42 cogniarch

Live Dashboard

The real-time dashboard provides 5 visualization lenses, each showing a different aspect of the simulation:

cogniarch --live --live-port=8001 --agents=8 --tom --coalitions --language --evolution

Lens	View	Key Bindings
Physical	Agent positions, resources, transmission lines	Default
Social	Trust networks, coalition hulls, relationship inspector	R toggle all trust lines, C toggle coalition hulls
Cognitive	SRIE cascade, metacognition state, ToM prediction arrows, calibration curves	D toggle deliberation heatmap, M toggle ToM lines
Cultural	Cultural group territories, language bubbles, transmission events, variant frequencies	L toggle language bubbles, T toggle transmission lines
Temporal	Agent movement trails, metric sparklines, action history, emergence events	T toggle trails, A toggle all/selected trails

Switch lenses with the toolbar buttons or keys 1-5. Click any agent to inspect its detailed state in the side panel.

LLM Setup

The LLMStrategy delegates agent decisions to a language model via any OpenAI-compatible endpoint. This works with cloud APIs, local models, and everything in between.

1. Copy the example config:

cp .env.example .env

2. Set your endpoint and API key in .env:

Provider	AUTOCOG_LLM_BASE_URL	API key needed?
Anthropic	https://api.anthropic.com/v1	Yes
OpenAI	https://api.openai.com/v1	Yes
Ollama	http://localhost:11434/v1	No
vLLM	http://localhost:8000/v1	No
LM Studio	http://localhost:1234/v1	No
Together AI	https://api.together.xyz/v1	Yes
Groq	https://api.groq.com/openai/v1	Yes
OpenRouter	https://openrouter.ai/api/v1	Yes

3. Set model names to match your provider:

AUTOCOG_LLM_MODEL=llama3.2          # Ollama
AUTOCOG_LLM_MODEL=gpt-4o            # OpenAI
AUTOCOG_LLM_MODEL=opus              # Anthropic

4. Run with LLM enabled:

cogniarch --llm

Or pass everything via CLI:

cogniarch --llm --llm-url=http://localhost:11434/v1 --model=llama3.2

Running Experiments

Define experiments in YAML with base config, conditions, and replicates:

name: "Architecture Comparison"
base:
  world_width: 28
  max_ticks: 350
  num_agents: 6

conditions:
  - name: "reactive"
    overrides:
      default_architecture: "reactive"
  - name: "dual_process"
    overrides:
      default_architecture: "dual_process"

replicates: 4
seed_start: 100
metrics: [survival_rate, cooperation_ratio, deliberation_rate]
output_dir: "data/results"

cogniarch experiment examples/architecture_comparison.yaml

Examples

File	Description
examples/reproduce_paper.ipynb	Reproduce our ALIFE paper findings in 5 minutes
examples/quickstart.py	Run your first simulation
examples/custom_archetype.py	Create custom personality profiles
examples/run_experiment.py	Batch experiments from code or YAML
examples/architecture_comparison.yaml	Compare cognitive architectures across replicates
examples/trait_evolution_study.yaml	Study trait drift over time
examples/social_dynamics_experiment.yaml	Coalition formation dynamics

Extending cogniarch

Custom Archetype

from src.agents.archetypes import ARCHETYPES
from src.agents.identity import PersonalityTraits

ARCHETYPES["scientist"] = {
    "description": "Explores methodically, shares discoveries",
    "traits": PersonalityTraits(
        cooperation_tendency=0.7, curiosity=0.85, risk_tolerance=0.6,
        resource_sharing=0.6, aggression=0.15, sociability=0.6,
    ),
    "color": "blue",
    "symbol": "Sc",
}

Custom Decision Strategy

Implement the DecisionStrategy protocol:

from src.awareness.types import Sensation, Reflection, Intention, Expression

class MyStrategy:
    def form_intention(self, sensation: Sensation, reflection: Reflection) -> Intention:
        ...  # Your decision logic

    def express(self, sensation: Sensation, reflection: Reflection, intention: Intention) -> Expression:
        ...  # Convert intention to action

Research

cogniarch was built to support empirical research on cognitive architecture effects in multi-agent systems. The paper "Aggression Monoculture" (submitted to ALIFE 2026) uses cogniarch to demonstrate how architectural choices -- not just agent personalities -- shape population-level behavioral dynamics.

• 56,000+ simulation runs across architecture and personality configurations
• Full trajectory datasets (SRIE state per agent per tick) for reproducible analysis
• Controlled experimental design via YAML batch runner with seed management
• Benchmark results -- Architecture rankings across 10 scenarios
• HuggingFace dataset -- Full 56K-run dataset with Parquet exports

Project Structure

src/
  awareness/          # SRIE pipeline: sensation, reflection, deliberation, protocols
  cognition/          # Decision strategies: hardcoded, personality, LLM, planning, ToM
  agents/             # Identity, archetypes, personality traits, evolution, registry
  simulation/         # Engine, world, actions, entities, renderer
  communication/      # Message protocol, mailbox, channel bus, emergent language
  memory/             # Episodic + social memory systems
  emergence/          # Pattern detection + metrics collection
  trajectory/         # Recording, export (JSONL/CSV/Parquet), loading, DuckDB catalog
  visualization/      # Live WebSocket dashboard, 5 lenses, static assets
  social/             # Coalition formation, coordination, dissolution
  evolution/          # Genetics, reproduction, cultural transmission, lineage
  theory_of_mind/     # Mental modeling, intention prediction, strategic reasoning
  metacognition/      # Self-model, FOK monitoring, metacognitive control
  planning/           # Goal decomposition, plan execution, progress monitoring
  persistence/        # Save/load simulation checkpoints
  experiments/        # Experiment runner, analysis, provenance tracking
  config.py           # Pydantic Settings (AUTOCOG_* env vars)
  main.py             # CLI entry point

License

MIT