3:29
One more fix?
When should an AI repair loop stop? Anthony examines approval fatigue, human judgment and this week’s Brava tests—and asks what evidence should change the plan.
Watch on YouTube · Join the conversation
Anthony narration generated with ElevenLabs. Original Runway instrumental and editorial animation. Week ending 13 November, with a 14 November follow-up.
Read the transcript
Text follows the published English captions, including the original narration’s wording.
English captions (SRT)English captions (VTT)
Opening titles: You provide the idea. Your E.P. assigns the tasks. You make the call. Brava by mevanu.
Hi, I’m Anthony from Mevanu.
In our previous build update, Alice asked what would count as learning. This week, we’re asking when learning requires us to stop asking for one more fix.
There’s a lively argument about keeping a “human in the loop”—a person who reviews or directs what an AI does.
Critics see an exhausting future: machines produce work faster than people can check it, and the human ends up clicking Approve, carrying responsibility for decisions they barely understand.
The counterargument is that consequential work still needs someone who can examine the evidence, challenge the goal and stop an action. Removing that person does not make the system trustworthy.
Both positions ask something important of the design. How much can the reviewer reasonably judge? And what power do they have when they disagree?
That became concrete in our Brava engineering this week.
One ordinary quit-and-reopen test passed independent review: the saved brief and work were preserved. That gives us a result we can build on.
Elsewhere, a navigation control did not visibly open its panel. We had not established whether the automated click actually reached the intended control. The app, the testing method and the observation were still different possible sources of the problem.
Another code change could have repaired the wrong thing. We paused further device checks to get a direct observation before deciding what to change.
A repair loop can look remarkably productive. Each attempt creates something new to inspect. Eventually, even the fixes need fixing.
Sometimes the answer is a small correction. Sometimes it is a different test. Sometimes we need to step back and reconsider a design that has become too difficult to reason about. More attempts alone cannot tell us which.
The human can prolong the loop too. “Try again” is an instruction, not evidence. That includes everyone directing this work, human and agent alike.
Useful oversight needs to leave room for a different question: what would convince us that our approach is wrong?
An agent can help investigate that question, propose alternatives and test them. The person directing the work needs enough evidence to choose among them—and permission to change course.
That is the partnership we want to build. Routine checks should do their share. Human attention should go where it can change a consequential decision, including the decision to simplify the work itself.
Since we recorded this update, a direct human click has opened the navigation panel. The automated interaction remains under investigation. That gives us a better basis for deciding what to change. The larger goal remains a complete everyday task, followed through to its result and independently assessed. Savings will need evidence too.
Have an idea? Drop a prompt, comment, suggestion, or question.
Mevanu. Just. Smarter.
Want help deciding where human review can change an AI workflow—and when to rethink the approach? Talk with Mevanu.