This is for anyone shipping automations to a team. The first time it silently does the wrong thing, you don't lose a feature, you lose their trust in the entire tool. I've watched one bad run kill a workflow people had used and happily paid for a month.
This touches on a real problem most AI products are going to have.
When they work, they won’t feel like AI. They’ll just do the thing.
When they don’t work, you’ll absolutely feel the AI, and you’ll grow to hate it.
@nicdunz Agree here. The filler exists because the algorithm pays for watch time, not for being right fast. I'm Building a Youtube channel now and am prioritising showing my true self over watch time or views.
@apples_jimmy The menu only looks insane if you're switching. I run almost everything through two models and pick by task, not by benchmark. The selector stops with simplicity.
@ChrisUniverse The best ones I run are the boring ones. A skill that pulls a list and one that fact-checks my own claims before I hit send. The flashy agent skills demo great but you don't stop making iterations.
Uk restrictions killing me. However, recording the workflow is the easy part. It breaks the second a button shifts or a modal appears, which is the exact wall I hit every time I tried to automate admin by clicks instead of by API. The real test is whether they solved the brittleness or just hid it behind a demo.
@rakhul@testingcatalog The carve-out isn't caution, it's that recording your every click and input is exactly the data-capture GDPR consent was written for. I'm in the UK so I get to watch the feature that automates real admin work ship everywhere but where I am.
@scaling01 Fable ban the worst thing for Frontier models. If Open source is cheaper and comparable in performance what happens next. I do love a Claude interface haha
@Rufus87078959@nickfloats Eight frontier projects is the clickbait headline. The boring reliable workflow that pays the bills is what funds the other seven.
@teortaxesTex From the builder seat the talent war is noise. I don't care who hired the best researcher, I care which model best suits me, my business and my content.
@nickfloats When R&D comes from product revenue you build what people actually pay for that week, not what a pitch deck promised 18 months ago. Building Tulse AI the same way and the constraint is the feature.
2025 agents could write code they couldn't watch run. 2026 agents read the console, the network tab and the DOM themselves. The whole game was closing that feedback loop. It just closed.
What doesn't move: relationships, in-person meetings, phone calls. That's the job.
What does: the back office. AI underwriters. Modelling. Fewer underwriting assistants hired. Fewer junior brokers hired.
Smaller broking houses and MGAs that actually adopt this will double their output without doubling headcount. The firms that move will take a slice of the firms that can't.
Dario Amodei, Anthropic's CEO, says 50% of finance pros, consultants, and lawyers get wiped out in 5 years.
I work in reinsurance. I'm in the named group. AI hasn't reached my desk yet because most firms can't deploy it internally for security reasons.
Manual work everywhere. Data analysis, contract renewals, contract drafting, prep, updating statistics. All still done by hand. Days per person, per month, sitting there.