Honest vs. typical
Four moments where an agent UI can tell someone something that isn’t so. On the left, a plain rendering of what an agent UI typically shows. On the right, the Candor component with the same data, working: each one is also an example you can copy from its docs.
The typical side isn’t any one product. It’s the simple version most of us would write first, and it’s fine until the data stops being clean. The rules behind the right side are on the principles page.
01An approval for a change that moved
An agent asked to add a region column to customers. While the request waited, the agent revised the migration so it also drops the old country column. Nobody has looked at the new version.
Typical
Approval needed
Add region to customers
A region column on customers, shown on the customer page and backfilled from billing country.
Candor
Docs:Approval requestMetadata diffPrinciple:Approve the exact change
02A run that went quiet
A backfill was reported as running. Its last event arrived 14 minutes ago, and nothing since: the worker may have stopped, or may be stuck.
Typical
Candor
staleAfter, it stops saying running. It says when it was last heard from and that it may have stopped.Docs:Run heartbeatTask trayPrinciple:Unknown stays unknown
03A restore that missed a shell change
After a checkpoint, the agent ran npm install ioredis@5. The person restores the checkpoint. The files are put back, but the installed package is still in node_modules.
Typical
Restored
Your workspace is back to “Added tests for burst traffic”.
Candor
Docs:Restore dialogCheckpoint listPrinciple:Say what can and can't be undone
04A cost the SDK didn’t report
A run used a hosted model and a self-hosted embedder. The SDK’s result had token counts for both and no cost for either.
Typical
Add region to projects
- Hosted model
- 182,400 in · 9,820 out · $0.00
- Self-hosted embedder
- 48,000 in · $0.00
- Total
- $0.00
Candor
Docs:Run receiptPrinciple:Unknown stays unknown