AI · Overview
Agents that answer for themselves.
I build AI agents that touch money. That is the hardest place to put one, and the best place to learn what safety and security actually require.
What I build
I build agents that act for real businesses: a merchant agent that takes orders and bookings, an email agent that works a real mailbox, a voice agent that answers the phone. Every action that moves money or leaves the building passes a human checkpoint, and every call is written to a record that names who decided and why.
They are built on Claude, in Python and TypeScript, alongside the coding agents that maintain the systems themselves. I work inside Claude Code every day, and everything on this site was built that way.
Each one has a hard stop where a person takes over, and each one fails closed when it is missing a credential, a list, or an approval.
What I believe
AI recommends. People decide.
An agent can draft, score, and propose. A human owns the decision, and the system keeps the record of who owned it. I will not ship a system that hides the person behind the machine.
Say null where you have not measured.
A score with no observations behind it is not a score. It is a lie with a number on it. The rule in my code base is simple: no fabricated defaults, no made-up success rates, no grade for an agent nobody has watched.
Fail closed when the money is real.
If a sanctions list is empty, the screener must refuse, not wave the payment through. Every control that protects money has to fail safe, and every fail-open path is a bug.
Audit your own system like an adversary.
In 2026 I commissioned read-only audits of my own platform with one question: can it defend its numbers to an underwriter, an auditor, or a regulator? The honest answer was no, with file and line. That report is why the rules above exist.
An agent's agency is a security surface.
Agents hold credentials, send mail, and move funds. I treat every permission as a liability: least privilege by default, approval ladders for anything that leaves the building, and a kill switch a person can reach.
The community
Where this work actually happens. Direct links, no middlemen.
Anthropic
- Alignment
Alignment Science blog · research notes from the safety teams
- Red team
Frontier Red Team blog · cyber, bio, and autonomy evaluations
- Interp
Transformer Circuits · mechanistic interpretability papers
- Welfare
Exploring model welfare · the research program and its uncertainty
- Fellows
Anthropic Fellows Program · four months of mentored empirical research
Research organizations
- Evals
- Alignment
- Welfare
Eleos AI Research · Taking AI Welfare Seriously (Long, Sebo et al., 2024)
- Training
Discussion
- Forums
- Careers
Why it matters to me
I spent twenty years launching products inside regulated financial institutions, where every system had to explain itself to a risk committee before it reached a customer. AI agents deserve the same bar, and today most do not meet it.
The work I want to do next is the work between those two worlds: measuring what agents actually do, proving it, and building the controls that let a careful institution say yes.