The software industry has spent decades teaching machines to move faster. David Maulick is interested in the moment just before they move. It is the point where a vague request becomes a choice, where an engineering team discovers that six people have been politely agreeing to six different products, and where speed can turn ambiguity into a beautifully compiled mistake.
That territory is the premise of Scott AI, the company Maulick co-founded with Devin Cintron in 2025. Scott is a planning workspace for teams building with coding agents. Several agents can explore a problem in parallel, compare approaches and expose the places where their reasoning diverges. People then resolve the important tradeoffs and carry a clearer specification into tools such as Claude, Codex or Cursor.
It is an unusually social thesis for an AI company. The product begins with a roomful of disagreement. Its job is not to eliminate the humans from that room, but to make their choices visible before an execution agent races away with the wrong interpretation. Maulick's stated bet is that humans remain at the center of software development even as agents take on more of the labor.
Chapter oneThe engineer behind the engineers
Before Maulick built a tool about alignment, he worked on the kind of infrastructure that reveals why alignment matters. He joined Coinbase as a junior developer and left five and a half years later as a staff software engineer and engineering manager. His account of the period reads like a tour through the hidden machinery of a large technology company.
He helped rewrite the Coinbase app in React Native ahead of the company's direct public offering. He helped scale its design system until it powered 95 percent of Coinbase UI. He led the introduction of microservices into the GraphQL layer across Coinbase products. His Y Combinator biography adds another measure: platforms he helped build and lead handled 65 percent of Coinbase edge traffic and were used by more than 700 engineers each day.
These are multiplication problems disguised as software projects. An internal platform succeeds when someone else can move faster without needing to know every compromise beneath the floorboards. The work requires code, certainly, but also persuasion, documentation, defaults and a sharp sense of which decisions should be centralized. A platform engineer learns that technical consistency and human coordination are roommates. They may quarrel, but neither can afford the rent alone.
When Maulick announced his departure, he thanked the teammates and mentors who had trusted him with work that stretched him as both engineer and leader. The photograph accompanying the post was less ceremonial: laptop open, earbud in, ocean behind him. A corporate farewell delivered from a deck chair is an efficient little comedy about modern work. The scenery says holiday. The charger says otherwise.
“When I joined Coinbase I was a Junior developer. Fast forward to today, and I’ve had the privilege of contributing to some of Coinbase’s most ambitious engineering projects.”David Maulick, on leaving Coinbase
Earlier evidenceA thesis about tools people misuse
The prehistory of Scott AI appears in a different technical problem. In 2018, Maulick completed a master's degree in electrical engineering at California Polytechnic State University, San Luis Obispo. His thesis compared the usability of security mechanisms provided by iOS and Android. It asked why developers so often misused the protections their platforms supplied.
Maulick proposed an evaluation framework and applied it to seven security mechanisms. He found that both mobile platforms adequately supported popular functions such as key storage and HTTPS, while lower-level mechanisms such as encryption and message authentication codes suffered from old documentation, awkward APIs or insufficient abstraction. Both platforms offered little guidance for verifying implementations through testing.
Security and agentic software are different fields, but the usability question travels well. A capability can be technically powerful and practically clumsy. The gap between what a system permits and what a developer intends is where defects acquire excellent disguises. Maulick's work has repeatedly occupied that gap: first studying it, then building internal platforms around it, and now founding a company inside it.
The Scott planning loop
The new companySpeed discovers a planning problem
Maulick and Cintron had known each other for more than five years when they started Scott. Their backgrounds meet neatly at the product: Maulick brought the experience of developer platforms at Coinbase; Cintron had been a founding engineer and head of mobile at Comun. They entered Y Combinator's Fall 2025 batch, found a HackerHaus office and teased a project intended to help developers stay in flow.
When Scott emerged from stealth, the pitch was direct. Coding agents can build quickly, but they frequently over-architect, select obscure libraries or optimize for futures that never arrive. A faster agent does not repair a fuzzy requirement. It simply gives the requirement more momentum. Scott aims to intervene during specification discovery, when a team can still compare plans without paying the cost of implementing all of them.
The product was initially described as a Mac app compatible with different code-generation tools. That neutrality matters to the idea. Execution agents are naturally biased toward their own environments and methods; Scott positions itself as a decision layer above them. Its agents debate architecture, identify meaningful divergences and produce a portable plan. The human role becomes editor, judge and keeper of local context.
“We are laser focused on the spec discovery/ planning phase of working with coding agents.”David Maulick, explaining Scott AI's scope
Scott's current language has broadened from plan mode to a workspace where product teams capture context, align on specifications and send precise work to agents without overbuilding. The shift preserves the founding concern. Code generation is an execution layer. The difficult decision often lives upstream, in the conversation about what deserves to exist and which constraints are real.
A changing craftQuestioning yesterday's good abstractions
Maulick's recruiting note for Scott offers a revealing tension. At Coinbase, his old team maintained a design system that reached nearly every corner of the company's interface. Now he wonders whether whole teams will continue maintaining that layer when a Figma connection can bridge design and code directly. The observation carries more weight because it questions a category of work he once helped scale.
This is not a declaration that design systems have vanished. It is a founder examining which assumptions survive when implementation becomes cheaper. Large companies once standardized on a narrow set of languages partly because supporting more languages was expensive. If agents absorb more of that support work, perhaps teams can choose the language that best suits the task. Every old best practice receives a fresh hearing.
Maulick is equally skeptical of easy AI metrics. Scott's hiring page calls token usage a poor success measure and asks candidates how they reduced tokens for a task while preserving the productivity gain. On X, he has complained that agents generate too many code comments, creating stale explanations that can confuse future agents and waste more tokens. The view is practical: more machine activity is not automatically more useful work.
Completes an electrical engineering master's thesis on usable mobile security mechanisms.
Moves from junior developer to staff engineer and engineering manager during 5.5 years at Coinbase.
Starts Scott AI with Devin Cintron and joins Y Combinator.
Builds a planning workspace for product teams using multiple coding agents.
The betKeep the room small and the humans central
The company's philosophy is not only visible in the software. Scott says it wants a small, close-knit team in which everyone works on the company's most important problems. Its preferred noun is “builders,” a little broader than software engineers and a little less polite about departmental boundaries. The team expects people to scale themselves with agents while remaining accountable for the result.
There is an appealing restraint in this model of automation. Agents can widen the field of options, inspect a codebase and articulate disagreements at machine speed. They cannot own a team's purpose. Scott leaves the consequential selections with people and tries to give those people a better surface on which to choose. Human judgment is not ornamental. It is the scarce input around which the system is organized.
Maulick's career makes the bet legible. At Coinbase, he helped create leverage for hundreds of engineers. At Cal Poly, he studied how implementation tools could fail the people using them. With Scott, he is applying both instincts to a peculiar moment: software can arrive faster than a team can develop a shared understanding of it.
The coding agent looks eager. The cursor is blinking. Somewhere, six engineers are still deciding what “simple” means. David Maulick would like them to settle that first.