It’s using the ReAct pattern, which is straightforward and surprisingly effective, as previously discussed. This pattern gets the AI to respond by making statements in a Thought/Action/PAUSE/Observation loop:
• Thought: the AI has to state the current situation and what it needs to do next
• Action: here the AI can act, using a structured command,_ perhaps by speaking to the user or running a command from a list of available tools.
• PAUSE: at this point we stop the AI auto-generating text, take whatever it said in the “Action”… then actually go and do it
• Observation: finally we feed back in the result of the action as an observation, and restart the loop. The AI resumes with a “Thought.”
Generally with the ReAct pattern the tools made available to the AI allow it to query Google, or look up an article in Wikipedia, or do a calculation. Using tools decreases the risk of hallucination and gives the AI access to accurate, up-to-date data.
This is an early form of an agent model.