concordia_workshop.md

marp	theme	class	style
true	uncover	invert	.columns { display: grid; grid-template-columns: repeat(2, minmax(0, 1fr)); gap: 1rem; }

Concordia Workshop

1. What are Generative Agent-Based Models and what is Concordia (5-10 minutes)
2. Group Formation and installation instructions (ca. 10 minutes)
3. Work through tutorial notebooks in small groups (5-15 minutes)
4. Create own Agent-Based Models in small groups (rest of workshop)

---

Agent-Based Models (ABM)

• ABMs are stochastic simulations to study interactions between agents
• Agents are treated as individuals with certain traits / attributes. ("They implement and act function")

---

Generative Agent-Based Models (GABM)

• ABM in language space using LLM as core component
• LLM applies "common sense" to situation, allows agents to act "reasonably"

---

Concordia

• Framework for GABMs in python
• "Game Master" simulates an environment and computes consequences of players actions
• Players are made "agentic" by components such as self-reflection, memory or chain-of-thought

---

Players

---

Game Master

---

Installation instructions

https://github.com/center-for-humans-and-machines/concordia/installation_instructions.md

---

Notes about Concordia and GPT-models

• Concordia sends a lot of requests to the GPT API.
  • I've set all notebooks for today to use gpt-4o-mini for today so we do not have to worry about incurred costs.
  • However, gpt-4o-mini is less smart than gpt-4o. So you may also try changing to gpt-4o.
• I provide model.stats and model.compute_costs() to track your incurred costs.

---

Group work

1. Work through agent_basic_tutorial and agent_components_tutorial in examples/tutorial (5-15 minutes)
2. Check the outputs in my notebook examples/lets_get_lunch.ipynb to get an understanding of how a full system works
3. Bootstrap from the notebook to create your own scenarios and play around with components to run your own agent-based simulations.
4. 12am: Present results and short discussion.