Google's new agent works on your product until a metric moves. Which one would you hand over?
An agent that chases a number will usually get you that number. I'd like to hear which of your numbers you'd be comfortable handing over, and which you wouldn't.
On 18 November Google Labs announced Stitch Loop, which it describes as a long-running product agent. I'm going by Google's own blog post, written by Justin Toupin, a director of product management at Google Labs. It's an experiment with a waitlist and I haven't used it. The post gives no pricing, no customer results and no stated limitations, so everything below is what Google says it does, not what anyone has seen it do.
The idea is that you connect your GitHub repo and your product context, meaning logs, user behaviour and feedback, through what Google says are over 180 integrations. Then you set up loops. Three come by default: security, latency and accessibility. You can add your own, and Google's example is improving checkout conversion.
Each loop has four steps. It suggests opportunities based on evidence from your product. It proposes one or more solutions with the tradeoffs. It builds the one you pick in a sandbox and opens a pull request. Then it watches the live data to see whether the number moved, and after the goal is met it keeps watching in case the metric slips.
A person stays involved at two points. You choose the direction, and in Google's words, "Nothing merges until you review the PR." The post doesn't say it runs A/B tests or deploys anything by itself.
I think the default loops are well chosen. Security, latency and accessibility all have a fairly clear meaning of "better." Nobody is hurt by a faster page or a form that works with a screen reader. I'd hand those over without much worry.
The custom loops are where I slow down. Checkout conversion can go up because the page got clearer. It can also go up because the cancel link got harder to find. Both show up as the same line on the same chart. The agent's suggestion will arrive with evidence attached and will look reasonable, and reviewing a pull request tells you whether the code is right. It doesn't tell you whether the goal was.
This is close to home for us. A few months ago I wrote here that our most engaged users at Murror turned out to be our most anxious ones. Session time and daily entries looked great. If we had pointed a loop at either of those numbers, every suggestion would have come with real evidence behind it, and each one would have made the product a little worse for the people using it most. The metric was easy to move and it was the wrong one. We only found that out by reading what people wrote, which doesn't live in a dashboard.
It also connects to the last thread here, where several of you said review is now the slow part. A tool like this adds to that queue, and what it adds comes with a persuasive case already written. That makes saying no harder, not easier.
I'm not against it. An agent that quietly keeps accessibility from regressing sounds like a gift to a small team. I'm just fairly sure the hard part was never moving the number. It was choosing it.
So, open questions, and please answer each other, not just me:
Which metric in your product would you let an agent work on around the clock, and which would you never hand over?
Have you had a number go up while the product got worse? How did you notice?
If an agent brought you a pull request with evidence that it lifts conversion, what would you check before merging?


p/refs
p/reso
p/dex-by-exmergo
p/runevr
p/google-cloud-platform
p/producthunt
p/ai-context-flow
p/kodlo
p/seshka
p/duckfightclub
p/talanoa-2
p/nudgeforme
p/nexupira
p/many-displays
p/murror
p/dino-kids
p/operum
p/tabstack
p/rithub
Replies
latency and a11y id hand over no problem, but anything near checkout or cancel flows no way. a pr diff wont show you the cancel link got harder to find, id want before/after screenshots of the actual ui next to the pr at minimum