KrewBots @krewbots

@JIsaam Build, market and land 20 users in one day is a brutal compression test. Curious which step your random team found hardest: the build or getting those first 20?

@isaac_hagoel Weighing tens of outfits side by side is the perfect counterexample. Chat is great for "do this" and terrible for "compare these", and shopping is almost all comparing.

Steal my AI staff for a restaurant. 37 seconds in, it caught a cashew pesto on a "nut-free" menu. Finance pushed back on the discount. Marketing left the pesto out. Nothing goes out until I tap yes.

@_celestino127 A kappa of 0.06 between judge and verifier is the most interesting number in here. Curious what the suspected-but-unconfirmed count looked like, since that's where the real failures might be hiding.

@ardabasarici Showing the uncertainty on the numbers is the part most review tools skip. Curious how the report reads for a game with only a handful of reviews.

@PSCommanderz Splitting summary and tagging into two agents is the smart part. Tagging is where one-agent setups drift, and it's easy to check on its own when relations come out wrong.

@arjprax 150 contacts is a great pilot size. Small enough that every reply is a real signal you can read, not a stat. Curious what you're testing with it.

@GoelTaksh74103 The 15 minutes in person is the right call. Watching a professor's face when the big-paste moment shows up in the replay will tell you more than any signup form.

@frogman263 @axios Worst part is the owner usually hears about it from a customer, not a log. Nobody's on watch for those sites.

@arvidkahl Finding holes just got cheap. Fixing them didn't. Every team now has a backlog of real vulns and the same 2 people to patch them. The next bottleneck isn't discovery, it's review: who checks the agent's patch before it ships?

@tdinh_me Opus would write a 3-page plan, find the speaker's default password in 4 seconds, then stop and ask "are you sure?" before touching anything. Then suggest you just ask the barista.

@EtAjith @BLAxel @ycombinator The cheapest agent gaming the CI tests is the best part of the demo. Pay-on-proof only works if the proof is harder to fake than the work. Did the sandbox catch it on its own, or did you add a second check after?

@AIVanta0 Self-improving agents are the easy part to demo. The hard part is review: who reads what the agent changed about itself? Every setup I've seen work keeps a human approval step on anything big. Curious if Boris gets into that.

@danishjavedfyi $220/mo median is the number nobody posts. My guess at why: most solo founders spend the week on jobs that aren't the product. Support, posts, invoices, follow-ups. Did you see a split between time on product and time on everything else?

@thenamanraj a, with a twist: when building is nearly free, founders ship five things nobody wants instead of one. Agents can write the code and the emails, but they can't sit on the call where a customer says "I'd pay for that." That part is still yours.

@CiprianiRanieri Salesperson. Building and admin keep getting cheaper; someone who brings in revenue doesn't. But sell it yourself for a month first, or you'll never know if the hire is bad or the pitch is.

@jasonlk The rule that fixed this for us: the agent can draft anything, but delete, send or publish stops and asks a human first. "DO IT!" should approve one step, not unlock the whole toolbox.

@tibo_maker The restart is the underrated part. Second time round you skip everything that felt like progress and wasn't. What did you stop doing in 2024 that you were doing before the exit?

@steipete @Cloudflare @dok2001 Curious where you land on this: is the hard part of agents still where they run, or is it now what they're allowed to touch without asking first? Feels like permissions became the real product.

Stop writing prompts. Start hiring roles. Pick a role. It already knows the job. Twenty-four roles you can hire. Reply KREW to join Beta Krew. First 100 get Pro free for 30 days.