www.businessinsider.com - Claude Code's creator says his setup involves thousands of AI sub-agents doing 'deeper work' overnight - 2026-05-13

|------ Anthropic engineer Boris Cherny says his coding setup now involves "a few thousand" AI agents working for him overnight.

Cherny, the creator of Claude Code, described his AI workflow during an interview with Sequoia Capital on May 4.

His answers - including how he mostly runs the agents via his phone - highlighted how some Silicon Valley engineers are beginning to use AI systems less like chatbots and more like always-on autonomous assistants.

Cherny said he relies heavily on two Claude Code features built for persistent automation: /loops and Routines.

He said users can schedule /loops locally via cron, while Routines run recurring tasks on a server, allowing engineers to keep agents working after their laptops are closed. ...
-------|

Stories like this are popping up nearly every day now.

Yes, of course. There are also "stories" nearly every day that don't make it into popular, mainstream press.


"A pair of humanoid robots crossed a major medical milestone this week after successfully performing gallbladder removals in pigs for the first time ... "

Intuitive Surgical's DaVinci robots (and their copycats) are old news and have been on the market since early 2000s; if you have seen phrase "minimally invasive surgery" it most likely refers to all kinds of robot-assisted medical procedures, from head to toes.
"In 2012, it was used in an estimated 200,000 surgeries, most commonly for hysterectomies and prostate removals."

And robotics may or may not include AI, but often confused with AI because most are usually consumer-facing, even if it's as simple as Roomba.


The Singularity is Nearer.

Yes, every day, by definition... but none of what you posted before is even remotely close to Singularity, or much beyond automation of previously "manual" tasks.

In fact, most of what was described in that 'prospectus' I could run much cheaper on a network years ago with a logic module, few scripts and not very sophisticated task scheduler, with or without ERP or "supervising" the process.

I didn't call it AI.
__________

Cost of AI token is NOT the cost per transaction. - see drudge.com

* 18% of organizations now orchestrating multiple agents across workflows, up from 9% in the previous period
* A $0.04 chat can become a $1.20 orchestration when it requires tool retrieval, planning and subagents
* AI coding costs are expected to surpass average developer salaries by 2028
* Despite enterprises spending an average of $11.5 million on AI in 2026, most struggle to demonstrate any clear return on investment.

* Recent tests have shown Chinese models to be much cheaper "per task" (not necessarily per token) and do almost as good a job as best US models, like Mythos, that run on more advanced hardware.


Nor does it matter how much money you make "per transaction" (however you want to define it), if you don't have enough of them (volume) to make a profit based on your costs structure - it's basic economics... hence, as I posted:

They are not necessarily losing money on every transaction because defining "transaction" in AI is very complex and involves many different, non-uniform costs.
However, they are losing a lot of money because the costs are huge, and they will definitely need a lot of subscription volume and high revenue growth to pay for their infrastructure, recurring expenses and continuous upgrades.

"Cost per transaction" is just one, poorly chosen, unit of many factors in trying to evaluate whether "Will the revenue generated by AI catch up to the CapEx for the hyperscalers?"
__________