At three in the morning, Akshay Budhkar is doing some of his best work. He is also asleep. The founder made the joke while introducing scheduled AI agents that prepare reports overnight, and it contains his entire product philosophy in miniature. Software becomes useful when it can be trusted with more than a reply. It needs an assignment, a clock, a place to put the result and some agreement about what finished means. Preferably, it should let the chief executive remain horizontal.
Budhkar is the co-founder and CEO of Spine AI, a seven-person San Francisco company from Y Combinator's Summer 2023 batch. Its current product, Medley, sits inside Claude Code and Codex and turns a requested outcome into a coordinated plan for multiple agents. There are targets, budgets, deadlines, handoffs and moments when a human must approve the next move. The chat box, once the whole stage, becomes only one prop.
The route to that idea has been unusually visible. Spine began with tools for adding AI copilots to software products. It moved toward business-aware data analysts, then iterative research, then an infinite visual canvas for working across models. Agents arrived to fill the canvas. Missions arrived to govern the agents. A less patient observer might call this a collection of pivots. A fairer reading is that Budhkar has kept following the same irritation as the technology beneath him changes: intelligence without a working environment is a clever guest with nowhere to sit.
The useful side of research
Budhkar studied engineering at the University of Waterloo and completed a master's degree in computer science at the University of Toronto, where he worked in the computational linguistics community. His published work included experiments with word embeddings, topic models and text generation. The through-line was less romantic than the phrase artificial intelligence often sounds. How can abstract methods hold up against stubborn, specific data?
A professional connection through the Vector Institute gave that question a commercial address. Georgian, the Canadian investment firm, had a natural language processing project that matched his graduate work. Budhkar joined as an intern. The internship became a full-time offer, and that offer competed with a place in a PhD program.
I still get to work on research while applying it to real-world problems in our portfolio.Akshay Budhkar on choosing Georgian over a PhD offer
He chose Georgian. It was a revealing decision, and not merely because it saved him from a dissertation bibliography. At Georgian he progressed from applied research scientist to AI technical lead. The job involved carrying research into the operating reality of software companies, where a model must meet product constraints, company data and the occasional spreadsheet with opinions of its own. During a portfolio secondment at WorkFusion, he also worked closer to the executive and product questions surrounding applied AI.
By the time he co-founded Spine AI with Ashwin Venkatesh Raman in 2023, Budhkar had accumulated a practical suspicion of demos. A demonstration proves that a model can perform a move. A product must survive the entire game: context, mistakes, corrections, permissions and the fact that a user may ask a vague question before lunch and still expect a useful answer by four.
Six names for one complaint
Spine's first public pitch offered infrastructure for product teams building copilots. The next versions pushed further up the stack. AI analysts could learn business context and answer questions over databases, documents and the web. Research became iterative so a user could change an outline or deepen one section without throwing away the rest. The canvas broke the linear thread into blocks that could branch, run in parallel and show exactly which context fed each step.
Each iteration added a noun from the world of work. Analyst. Report. Canvas. Team. Mission. Budhkar was gradually replacing the vocabulary of prompts with the vocabulary of organizations. It is a subtle shift. Prompts describe what someone says to a machine. Missions describe the outcome for which someone remains accountable.
His own founder diary makes the attraction obvious. In one post, Budhkar described a week in which he was head of sales, recruiter, growth lead and the customer success team. AI could help draft a job post or summarize a call, but he could not hand it a role for the next fortnight. Every hat remained on his head, looking jaunty and feeling heavy.
Each hat is a full-time job: recruiting, pipeline, retention, GTM, and there is exactly one of me.Budhkar on the founder's crowded hat rack
Medley is the answer he and the team are now testing. A user starts a mission from the coding environment. The system decomposes the requested outcome into a task graph, assigns agents and models, preserves the state passed between them, and provides a browser view of progress and handoffs. Sensitive actions still require approval. The human is not removed from the arrangement; the human is promoted, with all the calendar implications that word carries.
The space around the model
Budhkar's favorite technical idea is the harness: the machinery around a foundation model that decides how work is planned, routed, checked and resumed. Models can reason well on the next step, he argues, while ambitious outcomes are global. They depend on what happens before and after any one answer. This is why he talks about context surviving a handoff and review triggering a focused repair. The model is talented. The harness remembers where everyone put the keys.
Spine says a version of this architecture produced its reported results across four public benchmarks. The useful fact is not the victory lap but the choice of experiment: keep the underlying model available to others, then change the way work is coordinated. The company is making a product argument through an evaluation argument.
Same family of models, a different working arrangement
There is an appealing modesty buried inside this thesis. If the harness matters, the foundation model does not need to be treated as a lone genius. Different models can take different steps. A planner can revise its route as new information appears. A reviewer can send weak work back for repair. Existing capability may be less like a finished appliance and more like electricity before the building has been wired.
The risk is equally plain. More agents create more chances for cost, delay and confusion. A plan can become bureaucracy at machine speed. Visible intermediate work can clarify a decision, or it can give the user a very detailed tour of the wrong neighborhood. Medley's product burden is to prove that orchestration earns its complexity. Budhkar's public emphasis on budgets, evidence, permissions and stopping conditions suggests that he knows the paperwork is part of the invention.
Ambition, with a stop condition
Budhkar has started questioning one of startup culture's most upholstered commandments: choose one narrow customer profile and focus. His argument is conditional, not cavalier. Focus was partly a response to human cognitive bandwidth. Agents can operate in parallel, talk to one another and adapt interfaces to different users. If orchestration works, some old limits on organizational attention may loosen.
The provocation fits his career. He has repeatedly stood between a general capability and a specific company, translating research into something that can be operated. At Georgian, the bridge ran from a paper to portfolio work. At Spine, it runs from an increasingly capable model to a mission someone can watch, steer and approve. Budhkar has changed institutions, cities and product language, but he has remained on the bridge.
His aspiration is a future in which people manage teams of AI agents. That phrase can sound grand until he describes the ordinary founder tasks he wants to delegate: recruiting, prospect research, follow-ups, competitive monitoring. The future arrives wearing sensible shoes and asking where the latest spreadsheet lives.
This may be Budhkar's most interesting quality as a builder. He is willing to talk about ambition, even ending one argument with the line, “So what's left is ambition.” Yet his product ideas keep dragging ambition back toward contracts, evidence and review. The dream gets a deadline. The agents get a budget. The human gets an attention queue.
At three in the morning, then, the desired scene is not magic. It is operations. Work proceeds without its founder because someone has specified the outcome, arranged the handoffs and decided what requires a waking adult. Budhkar is betting that the next chapter of AI will be written less like a conversation and more like a company handbook. If he is right, the cleverest feature may be knowing when to knock on the bedroom door.