# Cua Builds Infrastructure for Computer-Use AI Agents and Opens Key Roles

**Author:** Zero G Talent
**Published:** 2026-11-20
**Category:** sales
**Tags:** yc-hiring, ai agents, developer tools, infrastructure, software engineering, sales

[Read on Zero G Talent](https://zerogtalent.com/blog/cua-builds-infrastructure-for-computer-use-ai-agents-and-opens-key-roles)

---

[Cua](https://zerogtalent.com/ai-companies/cua) is building the developer tools and cloud container infrastructure required to let general AI agents safely operate computers and applications like humans do. Based in Halifax, Nova Scotia, the seed-stage company emerged from Y Combinator's P25 batch with over 9,000 GitHub stars across its open-source repositories. To expand its platform from research prototypes to production-grade agent systems, the team is hiring roles including a [Founding Engineer, Infra & Agent Systems](https://zerogtalent.com/ai-jobs/cua/founding-engineer-infra-agent-systems-78130) and a [Founding Technical GTM Lead](https://zerogtalent.com/ai-jobs/cua/founding-technical-gtm-lead-109610).

## Scoping isolated compute for general AI agents

Cua provides an open-source framework, container execution platform, and benchmark ecosystem designed for long-running computer-use agents. Rather than running agents directly on a developer's host machine, Cua provides isolated cloud environments across macOS, Linux, Windows, and Android.

In a presentation at [AI Engineer](https://www.youtube.com/watch?v=ZSQb5fzRFPw), founder [Francesco Bonacci](https://zerogtalent.com/people/francesco-bonacci) contrasted traditional "Computer Use 1.0" loops—which take full-desktop screenshots and lock out user input—with Cua's background execution layer. Cua Driver allows AI agents to interact with individual application windows using accessibility tree parsing and background click injection via Apple's Skylight framework, leaving the user's primary desktop active. 

To evaluate agent capabilities, the team developed CuaBench, a suite featuring more than 130 verifiable tasks across 42 environments. In a benchmark built with Snorkel AI targeting electrical engineering workflows in KiCad software, top frontier models completed only six of 25 complex tasks, highlighting the open challenge of multi-step GUI navigation. Alongside open-source tooling, Cua monetizes its core infrastructure by providing specialized execution environments and synthetic trajectory data to frontier research labs.

## From open-source experiment to Y Combinator

[Francesco Bonacci](https://zerogtalent.com/people/francesco-bonacci), who previously worked at Microsoft on projects including Xbox and AI benchmarks like Windows Arena, began developing Cua's core execution layer after experimenting with early multimodal models. Former co-founder [Alessandro Puppo](https://zerogtalent.com/people/alessandro-puppo) previously contributed to Cua's early stage.

As explained on [GitHub](https://www.youtube.com/watch?v=JWmbmn0KDkQ)'s Open Source Friday series, Cua began as a command-line interface called Loom for managing macOS virtual machines. After a Show HN post on Hacker News generated over 1,000 GitHub stars within 48 hours, the company applied to Y Combinator and was accepted into the startup accelerator.

During an interview on [CodeRabbit](https://www.youtube.com/watch?v=cIwv_ET0g9M), Bonacci detailed Cua's shift toward multi-agent orchestration. By embedding Cua Driver inside coding and general-purpose agent frameworks like Hermes, developers can allow background agents to handle long-running, non-API-based workflows overnight without interrupting daily work.

## Optimizing reinforcement learning infrastructure

At the AI Engineer conference, Cua's team detailed how Cua Fleet addresses compute inefficiency in reinforcement learning (RL) training for computer-use agents. Standard RL loops lose significant GPU utilization while waiting for sandbox environments to reset or provision. Cua Fleet uses a demand-based auto-scaler to manage warm sandbox pools across macOS, Linux, Windows, and Android containers, decoupling sandbox startup latency from GPU token generation.

## Open roles at Cua

Cua is scaling its engineering, research, and go-to-market teams to move its open-source frameworks and container cloud into broader deployment:

- **Engineering:** The [Founding Engineer, Infra & Agent Systems](https://zerogtalent.com/ai-jobs/cua/founding-engineer-infra-agent-systems-78130) ($100,000 - $150,000 USD / year, San Francisco, CA) designs and scales cross-cloud container orchestration to execute real AI agents on isolated macOS and Linux environments.
- **Sales & GTM:** The [Founding Technical GTM Lead](https://zerogtalent.com/ai-jobs/cua/founding-technical-gtm-lead-109610) ($150,000 - $200,000 USD / year, Remote US / San Francisco, CA) serves as the first dedicated commercial hire, working with the founders to establish sales strategy, pricing, and developer outreach.
- **Research:** The [Research Intern (Summer 2026)](https://zerogtalent.com/ai-jobs/cua/research-intern-summer-2026-78742) ($8,000 - $9,250 USD / month, Remote US / San Francisco, CA) prototypes, tests, and benchmarks multimodal LLM-based agent systems and data pipelines.

| Role | Location | Work Mode | Salary |
| --- | --- | --- | --- |
| [Founding Technical GTM Lead](https://zerogtalent.com/ai-jobs/cua/founding-technical-gtm-lead-109610) | San Francisco, CA, US / Remote (US) | Remote | $150,000 - $200,000 USD / year |
| [Research Intern (Summer 2026)](https://zerogtalent.com/ai-jobs/cua/research-intern-summer-2026-78742) | San Francisco, CA, US / Remote (US) | Remote | $8,000 - $9,250 USD / month |
| [Founding Engineer, Infra & Agent Systems](https://zerogtalent.com/ai-jobs/cua/founding-engineer-infra-agent-systems-78130) | San Francisco, CA, US | Onsite | $100,000 - $150,000 USD / year |

[See all open roles at Cua](https://zerogtalent.com/ai-companies/cua)

---
*Source: [Zero G Talent](https://zerogtalent.com), the job board for space, defense, AI, robotics and frontier-tech companies.*