Meet Ron

Ron Kagan · Personal agents · Essay 3 of 4

Personal agents: the series

How knowledge passes between agents

The Jagged Frontier study raises a question about knowledge transfer: how do we carry its findings and limits from one assistant to the next?

How assistants shared this research

I asked Claude to research this paper while I worked with another assistant on other tasks. The findings were saved for the next assistant to read.

Select a step to see what was handed over.

  1. Assign research
  2. Save findings
  3. Read files
  4. Submit draft

Ron directs

Ron assigns the paper and reviews what the page should say. He and another assistant continue other work while Claude researches.

Claude researches

Claude checks the paper, records findings and caveats, and proposes the resource addition.

GitHub keeps the record

Committed files and the pull request preserve sources, findings and limits. This draft also records next steps for review.

Another assistant integrates

Another assistant reads the saved record, checks the published paper, and turns it into a page draft. Reading those files is an explicit step.

Ron decides

Ron reviews the draft and decides what to keep, revise or publish. A saved commit does not approve publication.

The shared repository carries saved work. It does not transfer the whole conversation.

How symlinks can share local context

General technique: A symlink is a local file-system pointer. It can give a shared context file the filename an assistant expects. When the assistant opens that link, it reads the common file.

Git handles a different step: commits record changes, and pushing and pulling share that recorded work between machines. A symlink does not sync files or transfer chat history.

This explains the technique; it does not document the exact setup used for this research.

Read the sources