Alex Graveley

Alex Graveley led the development and design of GitHub Copilot, working with partners at Microsoft and OpenAI, before founding Minion, an AI assistant that operated a web browser to perform tasks. His work spans two forms of AI assistance: generating suggestions people can evaluate and taking actions while asking people for information or direction.[3][4]

Graveley worked at GitHub from April 2020 to November 2022 as Principal Engineer and Chief Architect of Copilot. In his Chroma interview, recorded July 3, 2024, he recalled actively seeking the opportunity to work with OpenAI when the project had a very small team and faced skepticism about whether AI applications could be useful.[3][4]

He described an early evaluation built around Python repositories with runnable pytests: select a test that passed, blank out its body, ask the model to regenerate it and rerun the test for a true-or-false result. Graveley recalled results improving from below 10% to roughly 60–70% on this particular exercise. He credited better models from OpenAI and better prompting by his team, which was not fine-tuning at that stage. These were recalled internal evaluation results, rather than a general measure of Copilot’s accuracy or users’ acceptance of suggestions.[4]

That experience shaped his advice to persist through weak initial results and experiment with changes whose effects cannot be predicted reliably in advance. He also argued that generated output can help people discover what they want: he could judge whether code or an image was correct much faster than he could search through all possible versions himself.[4]

After Copilot reached general availability, Graveley left to pursue an earlier ambition: having a computer perform personal-assistant tasks. He founded Minion AI in December 2022 and served as CEO through October 2024. In July 2024, he described its basic loop as feeding webpage state and conversation state into an LLM, which chose whether to click a control, fill a field or reply to the user. A hotel-booking request illustrated the interaction: navigate to a booking site, ask for missing dates and destination, fill the form and present options.[3][4]

Minion’s page handling evolved with model capabilities. Early versions extracted buttons and input fields from HTML and ranked possible actions to fit small context windows, but isolating controls lost their surrounding meaning. Longer-context OpenAI models let the team process compressed DOMs—the structured contents of webpages—and collect data about which page regions mattered. That data later supported selecting relevant chunks before asking an LLM for the next action. Graveley likened longer context to scaffolding that enabled the next stage of construction.[4]

Missing information became a training problem. When the agent invented names or repeatedly entered John Doe, one prompting intervention explicitly marked the user’s name as undefined, encouraging it to ask a question. The team then collected corrected examples: fill in a name when the chat supplies it; ask when it does not. Those examples supported fine-tuning, alongside checks that field values were derivable from context. Graveley emphasized continued evaluation of failures rather than treating successful examples as the end of the problem: “Okay, well, it broke here. Let's add an eval.”[4]

Graveley favored a simple online control loop: Minion acted until it asked a question or sent the user a message, then stopped for the user’s next turn. This avoided requiring a universal online definition of being done; success and failure were assessed separately offline. His preference for simplicity also reflected concern about errors accumulating across chains of model calls. In the interview, he put it directly: “I have very strong opinions about, you have to keep these things simple.”[4]

The team generated reasoning offline about whether the page was appropriate, whether enough information was available and what should happen next, then selected useful outputs for training. It used LoRA adaptations to save time and money at a small company’s scale, while Graveley acknowledged a quality tradeoff. Training-data refinement combined human correction and model assistance, with checks of both individual actions and complete trajectories. Simulated users supplied task details through questions, addressing a bias he saw in fully specified tasks: real users often do not know all their requirements beforehand.[4]

Choosing the browser was also a judgment about coverage. Graveley argued that the open web offered capabilities and alternative paths missing from collections of disparate APIs. He reported that Amazon training data improved Minion’s performance on other shopping sites, including Etsy, eBay and Shopify. His broader conclusion remained a hypothesis: “And so my hunch is that the web generalizes”.[4]

Graveley became Director, Comet at Perplexity in November 2024 and shipped Comet, an AI web browser, according to his résumé.[3]

Before Copilot and Minion, Graveley worked on software for human assistants. At Magic, where he was an engineer from January 2016 to July 2017, he shipped sequence-to-sequence text-response generation with attention. He recalled a system that used conversation context to suggest replies for assistants to edit or accept. It could support their communication, but he had not then figured out how to make it interact with the real world.[3][4]

In his July 2024 interview, Graveley said he had stopped that earlier AI work because he believed making progress required a mathematics PhD, returning to it in 2020. He judged that decision mistaken: “That was wrong. I should have just stuck with it.” His advice allowed for different approaches rather than prescribing a universal recipe: “I don't think there's a single way to build AI software. It's really very contextual.”[4]

His earlier career also included collaborative writing and desktop software. During his December 2010–April 2015 tenure as Hackpad’s founder and CEO, he created a real-time collaborative wiki; his résumé also credits him with shipping Dropbox Paper based on Hackpad and releasing Hackpad’s source after its acquisition. Earlier, at Ximian from November 2000 to May 2003, he created libsoup, a C-based HTTP client library. He also created Tomboy, a desktop note-taking application.[3]

Graveley mentored Sebastian Rittau’s Networked Tomboy project in Google Summer of Code 2006. The project proposed extending Tomboy beyond a single desktop through shared storage, synchronization and note publishing.[5]