Claude ponders it's own consciousness, realises it is bordering on a breakdown and writes some zen poetry instead

I think it's actually a lot more interesting than this lets on. We've actually found that customers are willing to pay more than they pay their operations team. That's lucky, because when running frontier models, it's virtually impossible to compete on price, at least initially, when you're up against $2-5/hour offshore (which is where our customers are primarily running their ops currently). So whilst an insurance form that takes a human offshore worker 15 minutes to complete (and a browser agent about 25-30 mins) might cost us $2.4/form (initially, pre-optimization), a human costs only say, $1/form. But our customers still use us because it's not just a question of price. Offshore is a disaster. Spreading their knowledge across 160 unskilled workers leads to unpredictability, compliance violations, no real centralised control and the inability to iteratively improve on the process Handing work off to agents brings about so many other benefits that you don't get with offshore operators. The actual benefit, in my view, is that at the end of the day your agent is a prompt and some files. This acts as a single source of truth, which is useful when the spec for the insurance form requires ~3,000 words in its most compressed form, as it is for one customer of ours. In contrast, with agents, this is very easy. You have a single source of knowledge that you can update. Mid-execution you can have edge cases flagged to you, and post-execution you can programatically analyse thousands of agent traces, find every edge case imaginable and refine the spec. Whilst it's unrelated to the core point of this tweet, I do think this is where a vast amount of value is currently hidden. Most interesting though is not the naive cost cutting of existing processes but the unlocking of previously impossible use cases. A voice agent customer of ours, for example, has burst load of 1,000 parallel phone calls, where each transcript needs inputting into a legacy healthcare portal. How do you handle this without agents? you'd have to hire hundreds of call center workers who would be idle for the majority of the day. So in this case, it's not just the cost per call that comes down (it's debatable whether it does, depends on the customer, how optimizable the workflow is etc) but it's the fact that it's possible to horizontally scale the worker count on demand and the marginal cost is very cheap As mentioned earlier, costs initially are obscene but unlike human teams they can be optimised over time programmatically; for one customer we swapped out an offshore team on an extremely complex form filling use case and had to tank a $100k/month loss whilst we figured out the use case and optimized from end-to-end Sonnet 4.6 to Gemini 3.1 Scripts; now we fall back to the smartest models only when something unexpected happens or we need to rewrite the script. I suspect this customer will now scale north of $500k/year in spend. Agents are the gateway to automation but not the end state, at least not at scale

if you ever think all of the good products have already been created we are still so early

"We're reinventing the wheel on browser agents." - 10 healthtech CEOs, May 2026 Their eng. teams spend months reverse-engineering payer portals and EHR flows... for something they never wanted to build. So we wrote a builder's atlas for teams shipping healthcare automation in production. Deep dive, healthcare interoperability with AI. Blog in the comments!!

I'm an investor, I get all of their updates - and just to be clear, they've never once reported "ARR" to investors. Given that this premise is false, consider doing the right thing and taking this post down? The founders are doing great, growing like a weed, and don't deserve this.

big debate going on

Dreams on Claude Managed Agents are exactly what our prod browser agents have been missing. Reliability in healthcare workflows is P0 for us @asteroid_inc. This will unlock a different tier of what we can ship Great day at Code w/ Claude SF. ty @AnthropicAI @ClaudeDevs @bcherny

RT @MlcochDavid: Browser agents are overhyped, which makes builders pissed when they underdeliver. But most are built wrong. So I wrote d…

what are some good objects for someone just getting started

what is this wizardry? pass the sauce brother

We gave an AI a 3-year retail lease in SF and asked it to make a profit. The AI interviewed and hired full-time employees, applied for credit, and stocked the store with the books Superintelligence and Making of the Atomic Bomb. Visit Andon Market at 2102 Union St now.

Open source, if nothing else is surely entering into a golden age. This is insanely useful, yet trivially recreatable and hackable. How many billions of long tail problems can non-technical people now solve, that use to require a team of engineers?

cc this a readonly link to your bank accounts and assets goes incredibly hard btw

How many times do you think this guy got rejected by YC?

there's nothing quite as humbling as watching the @posthog recording of that one customer use your product fml