Two Agents, One GitHub Repo

The Mithlond research workflow originally used Google Drive as the handoff layer between two AI systems — Gemini would dump research into a Drive doc, I'd copy-paste it, Claude would read it. Three steps, two of them manual.

We cut that down. GitHub is now the sole sync surface. Gemini writes research findings directly to research/findings/ in the repo, updates a shared task tracker, commits, and pushes. Claude pulls at the start of a session, reviews the findings, flags what needs fact-checking, and promotes solid material to the blog or policy briefs. No Drive, no copy-paste, no intermediary.

The more interesting improvement was adding a GEMINI.md orientation file to the repo — the equivalent of what I use for Claude, but written for Gemini's reasoning patterns. It defines the two-agent architecture explicitly, specifies the finding format (each finding gets a confidence level, a source, and a confabulation risk flag), and includes a warning about the most common failure mode: Gemini generating plausible-sounding regulatory citation numbers that don't exist.

Confabulation at the source is the core risk in any multi-agent research workflow. The mitigation we landed on: Gemini labels its own uncertainty, Claude does a verification pass before anything goes into a public document, and I do a final spot-check on anything that will be cited in a policy context. Three layers, each catching a different class of error.

The result is faster, more transparent, and produces a full audit trail in git of what each system contributed. If a finding turns out to be wrong, we can trace where it came from.