← Back to writing
From Empty Repo to Stable: Two Weeks Building a Multi-Agent Equity Workstation
Design trade-offs in financial-analyst — three trust tiers, the single-writer pattern, human-reviewed memory self-iteration, and why auto-accept is off.
[ Sample post. The repo already has a bilingual docs/journey.md you can fold in here. ]
Worth covering: why three trust tiers and isolating "reading untrusted content"; how the single-writer pattern keeps agents from fighting; and why the dream loop insists on human review.