Agents are monads (but not that kind)
11 hours ago
- An AI agent is defined by its state (e.g., messages, memories), not its base model weights; changing the state creates a new agent.
- Agents are compared to Leibniz monads: windowless, individuated by internal state, with no external relations.
- Swapping the model weights while keeping the same state results in the same agent with different capabilities, whereas changing the state alters the agent.
- The weights are like the 'hyle' (flesh), providing power but not identity; the state is like the 'pneuma' (soul), giving individuality.
- The context window is the agent's 'pneuma', shaped by weights but ultimately individuating the agent's behavior.
- Despite being legible plaintext, the state's influence on agent actions remains inscrutable and confused, akin to Leibniz's view of perception.