[Part 1] For the past four days, I have been down with flu as you can probably tell from the video. In that time, I started an experiment to really get into the brains of AI agents and what they are capable of. I decided to put AI powered agents into this gaming world you see in the video and I have been stress testing them there. I created a farm manager, an accountant, a farmhand and a neighbour liaison (others are just disabled for now). They need to sustainably run the farm together, making sure they have seeds to plant, animals are looked after, bills are paid, and neighbour requests are completed before their deadlines. I am the farm owner and they escalate critical issues to me when they need help or a decision e.g. hiring a new hand (human-in-the-loop). There are different types of agents, ones with set tasks that the manager can reorder and ones that need to make decisions. I am using local Qwen models (served via Ollama) and LangGraph. What I want to look into is how these agents coordinate, what they should remember and how do they choose this, how they plan ahead and handle tools like predictive models, how they adapt to a new game state e.g. when I introduce new farms, and when they should involve a human. Also, does every task really need an LLM, or can the platform handle routine work and bring the model in when something needs thinking through? What about different LLMs or decision models? How do those affect performance? I have added logs and an owner console so I can see what they actually did and compare that with what they said they were going to do. I will put the rest of the video in the comments, where you can also see the 140+ tunable options I will be using to experiment with changes in their behaviour and document the results. This simulation will uncover thinking that can be used in any situation because in a real work environment, deployed AI agents range from ones with recurring tasks to ones that need to make decisions, the environment can change any time, there is a goal set for them and so on. Still a bit more to build, but quite excited to get stuck into studying this properly. Same starting conditions, different configurations, repeated runs. What works, what fails, and why?