AWS demonstrates training a smaller shopping agent

AWS’s example teaches a smaller shopping assistant from a larger model’s successful action sequences. This form of knowledge distillation uses the teacher’s text as training targets: each step pairs the conversation so far with the next reasoning and tool action. Tools do not run during this training. AWS publishes no per-task cost or latency figure; those require measurements with the intended traffic and workload.

The project guide separates a preliminary tool-choice check from the full shopping test. Run the preliminary check first: if the student does not outperform its untrained base there, the guide says the heavier evaluation is not worthwhile. The full test uses held-out problems. Base and student share an endpoint, with the student attached as an adapter; selecting an inference component chooses which answers. Reference-model comparisons use cached runs, not fresh calls.

The scoring code reveals what that preliminary check measures. For a reference containing tool calls, it compares a nonempty ordered list of tool names, tolerating formatting differences and ignoring parameters. That does not establish whether the arguments are correct or the shopping task succeeds. Separate metrics report stricter formatting and turn type. Before extracting calls, the parser discards text through the last closing reasoning tag, keeping reasoning mentions out of the scored action.