Research notes and memos on how AI changes the way firms decide, spend and take risk.
AI Engineering
Build logs from a small fleet of AI agents.
I run a small fleet of AI agents for my teaching, my research and this site. These posts describe jobs that actually run: what each one does, the check that proves it worked, and the rules that keep it safe.
Other sections
Payment rails, bank–fintech partnerships, stablecoins and licensing: who carries the risk when something fails.
Latest
The Mechanical Check Test
One question decides whether an agent may act alone: can a cheap mechanical check prove the result? The post shows the test inside a running daemon.
Review What Actually Runs
A nightly code-review agent picks its files with a formula you can print: how often the machine runs a file, times how long since anyone read it.
A Security Check With No Model in It
Six checks, once a month, read-only, with no model calls and no secret ever printed: the most reliable job in my fleet has no model in it.
Ship the Read-Only Report First
To automate your contacts and obligations, start with a job that cannot write anything, then add write access one tested phase at a time.