Machine Learning for Physical Systems · Lecture 11

Agent trace lab

Keys: P presenter mode, C input checks, V verifier

An agent is a loop: the language model writes a tool call as text, and our program runs the tool.

The model only writes text: our program reads it, checks it and runs the tool.

        1. Send the call the model wrote: what comes back?
        2. Switch on the input checks and send it again.
        3. Correct the call and answer in mm.
        4. Switch on the verifier: is your answer released?

        Our program reads the text between the tags; edit E_Pa.

        A number in millimetres.

        Our program refuses values outside plausible ranges.

        Code recomputes from the task before an answer goes out.

        Steps used
        Deflection from our programin metres, as the tool returns it

        More options
        Be the agent: the error message of the input check

        How our program reads, checks and decides
        How the runs were recorded and graded