Verybear 1.4: discovery inbox, staged rollout, Microsoft sign-in
The Verybear team
Changelog · March 12, 2025 · 2 min read
1.4 is focused on two things AI teams asked for repeatedly: a faster way to act on what Verybear discovers, and a safer way to roll changes out. Plus a sign-in option a lot of you have been waiting on.
Discovery inbox
New flows and optimizations used to live in a flat list. They now land in a proper inbox, grouped by type and sorted by how much traffic or spend they'd affect. A proposal like "Explain this codebase → move to Arch (local)" now shows its evidence inline (98% of prompts answered from local context, two weeks of shadow-run quality held steady), so approving it is a ten-second decision instead of a research project.
- New flow proposals now show the real requests that triggered them, not just a count.
- Optimizations show estimated monthly savings up front, before you open the detail view.
- Approve, dismiss, or snooze a proposal for 30 days. Snoozed items resurface automatically if the pattern keeps growing.
Staged rollout
Policy and flow changes can now target a single team before they sync company-wide. Publish a flow update to Engineering only, watch quality and spend for a day or a week, then promote it to every desk, or roll it back before anyone else sees it. This was the single most requested feature from Business and Enterprise customers, and it's live for both tiers today.
Microsoft sign-in
Microsoft joins Google and GitHub as a sign-in option, available on every tier, including Free. If your company runs on Microsoft 365, people can now join their first team without a new password to manage.
$ verybear status
✓ signed in as [email protected] (Ops)
✓ desk connected: Pi, Claude
✓ policy revision r184 (synced 2 min ago)
✓ 10 flows active, 1 draftAlso in this release
- Quality signal now shown per flow version, so you can see whether an edit helped before it fully rolls out.
- CLI: `verybear sync` now reports which policy revision it pulled, for easier debugging on shared machines.
- Fixed: budget alerts could double-fire when a team crossed 80% twice in the same day.
Keep reading
Why we route tasks, not models
Model routers pick a cheaper model for the same prompt. That's not the problem worth solving. The problem is that nobody wrote down how the task should be done in the first place.
Morgan Idris
January 14, 2025
Your AI bill is a judgment problem, not a token problem
Finance keeps asking which model to cut. Wrong question. The spend that actually hurts is the spend nobody decided to make.
Morgan Idris
January 28, 2025
