We play a game of probabilities
Picture AI as a character in a game you play in. Every answer it gives is a roll of the dice, weighted by the skills and items you grinded. Together they shape the probability landscape of outcomes. In AI jargon, this character is called an agent (an LLM plus its environment). Many people already run multiple agents, some working on their own and others when instructed.
The engine
The LLM, a large language model, is the agent's engine. It writes one word at a time, each word guided by probability. Anthropic's LLMs are Haiku, Sonnet, Opus and Fable, from the lightest engine to the most powerful. The trick is to shape the landscape an LLM traverses. The industry calls this context engineering.
Skills and items
In a game, the skills of your character grow while you play. An agent's skills grow through your corrections. For example, my writing skill began with a single rule and now lists every word I have banned, "rather than" among them. Bare LLMs reach for that phrase constantly. My agent has dropped it because its landscape makes the phrase improbable.
Items are everything else shaping the landscape of the agent. With your notes, it writes about your work in your words. A connector to your mailbox lets it send the mail it would otherwise only draft for you.
The landscape
The landscape is thus composed of skills and items. For example, my scout agent explores for new papers on its own. Another agent called extractor reads the articles the scout found. A third writes when I ask. A fourth checks every draft before it reaches me. You command the whole landscape. Agents cooperate, learn new skills and deliver items, adjusting the probability landscape with time.
Hidden stats
Some probabilities are hidden, yet they still shape the landscape. In a preprint, Cloud and colleagues (2025) took a model that liked owls and had it write nothing but numbers. A second model trained on those numbers came to like owls too, even with every mention of owls filtered out. This worked only when both models shared a base model. A trait can travel into a model inside data that looks clean. You read hidden stats by watching how your agent plays your real tasks.
Your first build
Open a project in ChatGPT or a similar app and load it with your notes and standing instructions.
Pick an engine that fits the task: a light model for quick replies, a powerful one for analysis and long texts. Write one skill for a task you repeat, such as the mail your team sends to clients. Fill the inventory with what a new colleague would need in their first week. Once the skill and the notes work, add a connector to mail or the calendar. Add a second agent when the first one has more work than it can do well.
The player
Playing with AI, you build agent after agent, the way a Civilization player starts empire after empire and a Hades player runs again after every death. Each agent multiplies what you know and teaches you how to think about the next one. Everyone who picks a model gets the same engine. But the inventory and the way you think are yours. They shape the landscape your agents play in.