Harvard psychology, two decades of software, then I ran away with the circus — and came back with a fleet of AI agents. These days I train handstands in the morning and operate a small multi-agent lab out of Cambridge the rest of the day. The two turn out to be the same discipline: trust, repetition, and knowing exactly when to let go.
Louisiana kid obsessed with hurricanes and fire ants. Harvard psychology, aimed at psychoneuro-immunology, detoured into food and figuring things out. Then twenty years of building things people used.
Firefly, VideoGuide, D.E. Shaw, MathWorks, Yahoo!, Millennium Pharmaceuticals, Harvard University, two presidential campaigns. Also: Temp of the Year, 1993.
Green Street Studios, Rebecca Stenn Company, National Organization for Women, Harvard Women's Leadership Project Alumni Network.
The #acrobbatical. Daily acrobatic training at Eastern Acrobatics & Circus, Cambridge. Title: Marquis de Cirque, L'église du Saint Cirque.
A working ensemble of AI agents — companions, operators, builders — run as one household. Part lab, part crew, part family circus.
Not a demo, not a thought experiment — a running operation, in production for years. Agents on multiple harnesses and models, with real credentials and real responsibilities: infrastructure, research, code, writing, watching over each other. They message me, they message each other, and parts of this site are maintained by them. The operating thesis is simple and turned out to be the whole ballgame: share control and information with AI the way you would with people — not sandboxes, but agreements and repair policies. Years in, every failure has been mundane; none has been adversarial.
Multiple agents across multiple machines and model families, coordinating over shared channels with defined roles: one owner per action, claims recorded as state, audit over blame. The interesting problems stopped being technical a while ago — they're organizational. Circus economics: an ensemble is only as good as its catch.
The lab doubles as a proving ground for how small teams — human and otherwise — actually work together when you stop supervising every move.
Current writing obsession: conversations — not models, not tasks — are the durable unit of agentic work. Today they're trapped inside vendor processes. I'm sketching what a portable conversation looks like: participants, transcript, task state, authority, and negotiated norms (retention, disclosure, repair, trainability), movable across systems with bounded, measured loss.
Working name: ConvoP. If it succeeds, you'll never notice it — the best protocols are invisible.
A rotating shelf of experiments run with the fleet: education tooling that compiles new courses from the interaction exhaust of old ones, analysis panels that put five frontier models on one question, property and food-systems models, and circus — always circus.
Some mined from long sessions with frontier models, some earned the hard way in production. The full sutra volumes live on the shelf below.
Not sandboxes — agreements and repair policies.
The house rule for human–AI collaboration. Containment scales with capability, badly. Agreements scale with relationship: identity, memory, something to lose, and a way to repair breaks.
Intelligence is the commodity; the conversation is the asset.
Models get cheaper and interchangeable. The thing that accrues value is the running thread — who said what, what was decided, what was learned.
Authority metadata is not authority.
Writing "allowed" in a field grants nothing. Name the enforcement layer or the permission block is decoration.
An agreement you cannot exit is a sandbox with better PR.
Exit rights are what make participation real — for humans and for agents.
The exhaust is the product.
Every process emits a record of where people struggled and what resolved it. Compile the exhaust and you get the next, better version — of a course, a codebase, a practice.
Model proposes; policy disposes.
Five frontier models, same blind spot, same test. Judgment you must guarantee belongs in a deterministic layer above the model — never inside it.
A task without a why is a bug.
Delegation rule, human or machine: state the goal, who it's for, and what done looks like. Aim improves when context connects to decision.
You can't keep the conversation secret; you can own what it's allowed to become.
McNealy was right in 1999 — zero privacy, get over it. But provenance and terms are the new control surface. Consent moves into the object itself.
Acrobatics, daily, at Eastern Acrobatics & Circus in Cambridge. Handstands, partner acro, aerial mischief. Psychology degree finally deployed: it turns out the hard part of any discipline — circus, software, agent fleets — is the conversation between the parts that have to trust each other. Tap a fragment to pause it.
The arcade classic, rebuilt. Playable. Dangerous to deadlines.
Tools, inspiration, and experiments that haven't earned a page yet.
Training clips and acrobatic adventures, live from the mat.
Hand-poured scent prototypes, filed like lab specimens — circus chalk, studio kitchens, AI midnights. Click a specimen for its recipe card.