Principles

I am building a company in San Francisco. Before this I was at Hexo Labs building agents that run the research loop themselves: SIA, a self-improving autoresearcher that rewrites its own scaffolds, hit 80.81% on GPQA Diamond with no human touching it between generations and is deployed at Lawrence Livermore National Laboratory. Before that, RL for computer-use agents at AGI Inc. The long bet is program synthesis.

The path here was not a straight line.

I turned down my first job offer because the interviews were too easy. It was the end of college, the Boston job paid exceptionally well, and I had no backup. I told my parents I would find something better. A few months later I joined a boutique consulting firm, and within months I was leading a team and reporting directly to clients like Bread Financial and California Lawyers for the Arts.

Then the firm gave me one month to move to Dallas. I was 22, new to the country, with no savings: a hundred dollars from my parents. Housing, cooking, licenses, I figured out alone.

Dallas was fine, and fine was not the goal. I picked the day of the Stanford Codex Hackathon as my deadline, packed everything into a U-Haul, and moved to San Francisco. I lived in hostels, worked East Coast hours, 5 AM to 4 PM, then went to every tech and startup event in the city.

I applied to Stanford's semester program. I was not going to spend two years on a Masters. I took High Performance Computing and Convex Optimization under Stephen Boyd, ran protein simulations on supercomputers, and solved optimization problems for rocket propulsion. On free days I snuck into MBA and astrophysics classes because they were interesting. My peers there were some of the sharpest and most ambitious people I had met.

Meanwhile my company got acquired and the work went stale. I needed to build. The startup route felt crowded and non-contrarian, so I took the other path: Buy then Build. Friends and family cautioned against it. I spent six months on due diligence, talking to franchise owners, local businessmen, and consultants, then signed the FDD on Mister Sparky, a residential electrician business in Palo Alto. I settled the lease, bought the equipment, hired a master electrician, and started the SBA loan.

Two months in, the numbers broke. Bay Area rates and licensing delays blew past the projected starting capital, and I had already put in 90% of my savings. I refused to cover the gap with borrowed money, so I cancelled the loan and walked away. I don't regret it. Given the same information, I would make the same call again.

I contracted for a couple of companies, including YellowPad, built by a Stanford Law graduate. At a hackathon, AGI Inc scouted me. I joined as a founding researcher under a Stanford PhD, implemented GRPO for training agents, and co-authored the REAL benchmark at NeurIPS 2025. Some of the smartest people I have met, and I have met a lot at this point.

Then Hexo Labs, and now this. Two papers from that stretch carry my name first: SIA (opens in a new tab), on agents that update their own harness and weights, and Vector-Bench (opens in a new tab) (co-first author), on whether models can surgically edit SVG code. The rest are on Scholar (opens in a new tab).

What we built at AGI Inc →

AGI-0: #1 on OSWorld

OSWorld is the benchmark for computer-use agents on real desktops. In October 2025 we beat it (opens in a new tab): AGI-0 reached 76.26% task success across 369 real-world tasks, the top computer-use agent at the time.

AGI-0 is an end-to-end trained computer-use policy. It acts directly on real VMs (Ubuntu, Windows, macOS) with low-level control only: clicks, typing, and hotkeys. Pixels in, actions out, no API shortcuts.

We trained it with GRPO in an online RL loop. Run the policy on real machines, score by execution with verifiers, update, repeat. The task pool was hundreds of thousands of synthetic and OSWorld tasks plus our internal REAL browser environments. The agent also self-verifies: it checks each outcome and, when a step fails, corrects on the next turn.

Built by Diego Caples (opens in a new tab), Atharva Gundawar (opens in a new tab), Prannay Hebbar (opens in a new tab), Greg Tarr (opens in a new tab), and Joseph Junzhe Zhu (opens in a new tab). Read the full write-up↗ (opens in a new tab)

REAL: the environment underneath it

Before the agent, the environment. REAL (opens in a new tab) (NeurIPS 2025) is a set of deterministic, high-fidelity replicas of 11 real websites with 112 benchmark tasks. Deterministic matters: the same action gets the same result every time, so you grade by execution instead of by rubric, and you can train on it, not just test on it. When we published, frontier models topped out around 41%.

The same stack shipped as agisdk (opens in a new tab): the web agent, the harness, sandbox-site access, and the benchmarking, in one Python package. 17k downloads.

theagi.company↗ (opens in a new tab)