FIG_AGT · An interactive simulation
Inside an AI Agent
An AI agent is more than an LLM call. Break the machine to see why. Run a request. Remove a component. Watch what fails.
What the machine shows
The model only proposes. It reads the cards it is given and prints a slip. It never opens the cabinet or turns the calculator crank. The runtime does that work, and decides whether to do it at all.
Knowledge comes from three places. The sealed drum is what the model learned in training. The memory drum is what this system saved about you. The cabinet holds your private documents. Only cards on the carriage reach the model.
A tool call is untrusted input. The schema gauge rejects malformed arguments, the permission bolt refuses tools this run must not use, and the timer stops a slow tool from holding the run.
Checks live outside the model. The verifier recomputes from the evidence. Turn it off and a wrong answer can leave the machine looking exactly like a right one.