Research talk

Evaluating document-agent workflows

A talk template for framing task, method, evidence, limitations, and future work.

Research question

Can agents revise rich documents without breaking portability or trust boundaries?

The unit of analysis is the package, not a chat transcript.

Method

Compare edits before and after validation.

Input Operation Check
Unpacked HTMLX Agent revises package files Validate directory
Packed `.htmlx` Runtime opens document Check rendering and resources

Limitations

Do not overclaim: this is a package workflow, not a universal document standard yet.

Use this final slide to state evidence gaps and next experiments.