Notebook Entry 014
Versioned Filenames Prevent Silent Overwrite Failures in Multi-Agent Systems
April 11, 2026
How do I stop silent overwrite failures when multiple AI tools touch the same files?
What happened
During a session reviewing Cowork's local folder inventory, it was discovered that Cowork had a known bug where saving a file with the same filename silently fails — the operation appears to succeed but the file is not actually updated. This means any save-progress operation using fixed filenames could appear to complete while actually writing nothing. Combined with the truncation failure pattern from entry 001, this creates a compounding risk: the file appears to exist, appears to have content, but may be stale or incomplete.
Root cause
Fixed filenames create a collision problem in multi-agent systems. When Claude Projects, Cowork, and Claude Code all have access to the same files with the same names, the last writer wins — but a silent write failure means no one wins and no one knows. The system has no mechanism to detect that the intended write didn't actually happen.
Decision made
Session-numbered filenames (CORE-01a-evergreen-patterns-s65.md) replace overwrites entirely. Rules: keep two most recent versions in the active folder (rollback safety net), weekly Cowork task moves older versions to REF-Past-Versions, read-before-write rule (always read the current highest-numbered version before writing a new one), line count verification (if new file < 80% of previous file's line count, flag as potential corruption — don't save silently), one writer rule (Claude Projects writes CORE files; Cowork and Code read only).
The lesson
Multi-agent systems need file naming conventions that make version history structural, not optional. When multiple AI tools touch the same files, you can't rely on any single tool's success signal — the system needs an architecture that makes silent failures visible. This is why version control exists in software development; AI implementation systems need the same discipline applied to their operating documents.
One implementation insight every other Tuesday.
What actually works when you ship AI in real organizations. Nothing you could get from a press release.