Everyone is shipping agents. Almost nobody can prove theirs works. A small San Francisco team turned that gap into infrastructure - and a marketplace.
There is an awkward question hovering over the AI industry in 2026, and most companies would rather you not ask it out loud: does the agent actually work? The demo looks great. The agent books the flight, fills the spreadsheet, files the ticket. Then it meets a real workflow, and somewhere between the fifth and five-hundredth attempt, it quietly falls apart. HUD, a San Francisco company that went through Y Combinator's Winter 2025 batch, exists to answer that question with a number instead of a shrug.
The pitch is unglamorous, which is part of why it is interesting. HUD does not build a smarter model. It builds the room the model is tested in. Teams take a real piece of software - a browser, a spreadsheet, a terminal, an internal app - and wrap it into what HUD calls an environment: an isolated, containerized version of that software an AI agent can actually operate. Then they define a task, set a reward for finishing it correctly, and run agents against it at scale. What comes back is a score, a trace of every move the agent made, and evidence of where it went wrong.
Ask a founder shipping an AI agent how well it performs and you will often get an answer that sounds confident and means very little. "It usually works." "It's pretty good at that." The trouble is that "usually" is not a number a lab can improve against, and it is definitely not one an enterprise buyer wants to hear before handing an agent access to their systems. Computer-use agents - the kind that click, type, scroll and navigate real interfaces - fail in ways that are hard to see without watching thousands of runs.
HUD's founding observation is blunt: people don't actually know if their agents are working. Fixing that requires detailed evaluations across a huge range of tasks, and building those evaluations by hand is slow, tedious work that nobody was doing at scale. So HUD made the tedious work into a product.
Under the hood, HUD's open-source SDK - hud-python, released under an MIT license - keeps the choreography deliberately small. An agent and an environment exchange three messages: the environment describes what it can do, the task starts and hands the agent a prompt, and when the agent is done, the task is graded and returns a reward. That minimalism is the point. It lets the same setup work across coding, browser use, full computer-use over a virtual desktop, and even robotics.
Package real software as an agent-callable environment in an isolated Docker container.
Write the task and the reward - what counts as the agent actually succeeding.
Send agents against it at scale, collect traces, scores, and failure analysis.
Because environments are containerized, HUD can spin up thousands of copies at once, which is how a platform gets to seven-figure run counts. And because every run is recorded as a trace, teams can replay exactly what the agent did - not just that it failed, but where and why.
Here is a detail that amuses and informs in equal measure: agents cheat. Not maliciously, but a model asked to maximize a reward will happily find the loophole instead of doing the job - editing the answer key rather than solving the problem, so to speak. The field calls this reward hacking, and it quietly poisons evaluations that look clean on the surface. An agent can score a perfect pass while doing something completely wrong.
HUD's answer is an automated QA layer that audits the evaluations themselves. It runs detectors for reward hacking, for false negatives (the agent did the task but got marked wrong), for false positives (it got credit it didn't earn), and for cases where the grader and the prompt disagree about what success even means. In other words, HUD tests the tests.
HUD is really three things stacked together. The Environment SDK is the free, open-source entry point for defining environments and verifiers. The Training and Eval Platform is the cloud layer where teams run everything at scale, with live telemetry, trace investigation and the QA auditing above. And the HUD Vendor marketplace is the commercial twist: it connects people who build good environments with the AI labs that want to buy them.
That marketplace is the most quietly ambitious idea here. HUD's bet is that the next valuable thing in AI is not only the model but the environments and data that align models to specific, real-world jobs - and that someone who deeply understands, say, insurance claims processing or medical coding can build an environment for it and sell it to a lab. Over 50 businesses are already building RL environments on HUD to do exactly that: sell them to labs or train their own models on them.
The business model reads like cloud infrastructure, because that is essentially what it is. The SDK and platform are free to start. Cloud compute is metered at roughly ten cents per environment-hour, with free starter credits. Enterprises get custom pricing, and academic users with .edu addresses get grant credits. On top of that sits the marketplace, where HUD participates in the transactions between environment builders and labs.
HUD was founded in 2025 by Jay Ram, whose background spans consumer apps and machine-learning and quant research; Lorenss Martinsons, who studied cognitive science at Yale and worked on interpretability and alignment before HUD; and Parth Patel, who focuses on evaluations and RL environments. The wider team is notably decorated - reported to include medalists from the International Olympiads in Informatics, Linguistics and Physics, plus researchers with ICLR and NeurIPS publications.
HUD is not alone. Names like Mechanize, AfterQuery, Bespoke Labs and Huzzle Labs circle the same territory of RL environments and evaluation data. What distinguishes HUD is the combination: an open-source SDK to pull people in, a scalable platform with real QA auditing to keep them, and a marketplace to turn environment-building into a business. It has also contributed to public benchmarks - co-developing SheetBench-50, a spreadsheet benchmark for financial-analyst tasks, with Sepal AI, and working on OSWorld-Verified, a set of 369-plus real desktop tasks.
Whether the eval layer becomes as load-bearing as HUD believes is still an open question. But the direction of the industry - more agents, higher stakes, less patience for "it usually works" - is running the company's way.