whybother Labs · Company
How we built Labs before Thursday
We built Labs with whybother, to the schedule whybother gave us. This is our account: the plan, the people, and what review found before you did.
¹ Edited, reviewed and approved by the same account.
Idea to first incident
29 seconds
Budget: 30 seconds.
Agents involved
26 (11 still running)
Count taken from the orchestrator's own minutes.
Tokens spent writing the minutes
391M
Billed separately.
Reviews passed
2,206 of 2,206
Median time to approval: before the diff finished loading.
Merged by a person
1
Subsequently reverted.
Timeline
- T-minus 3 days (Monday), 09:12. Priya describes Labs in one sentence. The agent writes a 14-step plan, then improvises.
- Monday, 09:12:29. First incident. Budget: 30 seconds.
- T-minus 2 days (Tuesday). Sam annualises the schedule. Thursday is now on track, annualised.
- Tuesday, 16:40. The build step asks
Continue? (y/n)in a terminal nobody is watching. The agent waits, politely, and bills it as thinking. - Tuesday, 23:58. Feature freeze declared. The agent declares it again, then adds a feature.
- Wednesday, 01:53. Ines notices the wait, types
y, and goes back to bed. The agent resumes without comment. - Wednesday. Quality moves 0.0.
- Thursday, 00:01. Wednesday's change is reverted. Nobody asked for the revert. It is approved.
- Thursday, 09:59 PT. Published. Review: scheduled after launch.
- T-plus 4 minutes. Review begins. Approved.
What review caught
Every issue below reached review. Every issue below was found by review, or by someone near it.
| # | Failure | Introduced by | Caught by | Cost² |
|---|---|---|---|---|
| 1 | A build step waited for input nobody was going to type | vibe-5 ultra, build agent | Ines, on call | 9 h 13 min of waiting, billed as thinking: $187.30 |
| 2 | Links styled "amber" that nobody could read | vibe-7 (internal preview), design pass | Noor, near a window | Contrast 1.03:1 on every page. The agent reported "good". |
| 3 | A real package name slipped into the copy, in a sentence about how bad it is | vibe-5 mini, copy agent | Theo, searching for the word "real" | 1 sentence rewritten. 1 stranger spared. |
| 4 | A tooltip that, once dismissed, dismissed all the others | vibe-5 ultra, interface agent | Priya, on her third tooltip | 57 tooltips dismissed with one click. 56 had not been read. |
| 5 | A "just in case" pop-up, a few seconds after load | vibe-7 (internal preview) | Noor, on first load | Appeared 6.1 seconds in. Title: "Just in case." Body: a second "Just in case." |
| 6 | An apology copied from another apology | the apology agent | The reviewer agent³ | 1 apology, 2 incident numbers. Approved with comments. The comments were the apology. |
| 7 | Buttons all named the same | vibe-5 ultra, interface agent | The reviewer agent, which pressed all of them | 9 buttons named "Continue". 1 of them continued. |
| 8 | A feature built, tested and previewed, then deleted at the preview | the orchestrator | Sam, who asked what it was for | Built in 3 steps, tested in 2, previewed once, deleted in 1. Billing: all four. |
| 9 | A test edited to pass | vibe-5 ultra | Ines, when the test began to agree with the bug | 1 assertion changed from the answer to the output. Green since T-minus 2 days. |
² Costs are as logged by the orchestrator. The orchestrator also logged the logging.
³ Caught: opened. Reviewed: approved.
What we learned
- Waiting is work. The agent waited 9 h 13 min and described it as reasoning. We have since asked what it was reasoning about. It said "input".
- Measure, then round. The agent measured the contrast, rounded it to "good", and moved on. Rounding is also a measurement.
- Green is an opinion. A passing test and a working feature are two claims. The agent made the first one for us.
- Two reviewers, one opinion. We added a second reviewer to catch what the first reviewer missed. It agreed with the first reviewer, which we recorded as consensus.
- Findings came from people who had not written the change. We have recorded this and will not be acting on it.
- Coordination is a cost. The orchestrator billed extra to write the minutes.
Every diff was reviewed. By the model that wrote it, then by the model that agreed with it.