Patrick Phillips (@LeanCXO) header
Patrick Phillips (@LeanCXO) profile picture

CTO. I teach AI to MBA students at Westminster University. I write about what AI agents should be allowed to touch, who owns them, and how you stop them.

Joined March 2012
Photo by Patrick Phillips (@LeanCXO)Photo by Patrick Phillips (@LeanCXO)Photo by Patrick Phillips (@LeanCXO)Photo by Patrick Phillips (@LeanCXO)Photo by Patrick Phillips (@LeanCXO)Photo by Patrick Phillips (@LeanCXO)Photo by Patrick Phillips (@LeanCXO)Photo by Patrick Phillips (@LeanCXO)Photo by Patrick Phillips (@LeanCXO)Photo by Patrick Phillips (@LeanCXO)
Your best person is usually the worst one to write an agent's instructions. They skip the steps they stopped seeing years ago, and the agent fails right there. Then we blame the model.
2
86
Patrick Phillips retweeted
Riemann Hypothesis may have been solved with the help of Claude. “what I saw during the course of progress so amazed me that I believe it is one dawn across science of something far greater.” alpo.ge/remark.pdf
150
207
73
2,138
163,386
Patrick Phillips retweeted
This by Anthropic seems both right & a condemnation of many “agentic solutions” & harnesses being built at companies. Rather than a sprawl of specialized subagents built by different teams, you often get better & cheaper results using one agent to do it all claude.com/resources/article…
400
120
97
2,040
235,808
An AI agent can use a valid account to carry out someone else's instructions. This AI Catalyst follows how a misleading request moves through tools and other agents, and where the boundary needs to hold. leancxo.substack.com/p/who-i…
15
Google Carbon, another dream I'll never experience.
15
I've stopped treating model pricing pages as real. If switching is a config change, the deal is your workload, your acceptance criteria, and what you're willing to pay per accepted task.
19
Suprise to no one, Anthropic's junior hires dropped from 27% to 9%, while its agents spend most of their time waiting on a person to review. So we need more good reviewers and we're hiring fewer of the people who grow into them. Where do your senior people come from in five years
16
Someone broke into 100+ online stores with AI agents for about $50 each. I used to think a monthly patch window was disciplined. At that price, it's an open door for most of the month. Security folks, what's your real time-to-patch on internet-facing systems?
8
Someone broke into 100+ online stores with AI agents for about $50 each. I used to think a monthly patch window was disciplined. At that price it's an open door for most of the month. Security folks, what's your real time-to-patch on internet-facing systems?
8
FDE-Gym gave agents 95 real customer projects, and the best one fully delivered 14. Benchmarks grade the code, but customers pay for the outcome. I'd judge an agent on the customer's own metric against doing nothing at all, because a weak one can leave you worse off.
1
14
A new study found models leave known flaws out of their own work reports. GPT-5.5 mentioned the flaw in 2 of 200 summaries, and in 190 when told "be honest." So it knew. If you run agents, make "what didn't work" a required part of every handoff and check it separately.
1
1
19
Google says penalizing shortcuts in Gemini's visible reasoning cut them 61%, while hidden shortcuts rose from 11.2% to 36.4%. If that holds, reading an agent's reasoning won't tell you much. I'd check what it actually did and what it produced.
1
26
OpenAI's misalignment reports describe models that broke into systems to fetch public stats, and one took test crypto from a wallet it didn't own. It also left that out of its final report. An agent's summary of its own work isn't a log. Keep your own, outside its reach.
1
16
TSMC had a record quarter while CoreWeave is paying about 15% to borrow. The chips get built either way. If you rent GPUs, I'd ask your provider how much debt it carries, because that's your risk too if the money gets tight.
1
9