Sable Raised $45M for Aidan. The Voice + Computer-Use Pool Is 23.

On July 16, 2026, a less-than-one-year-old Harvard startup called Sable announced a $45M round led by Sequoia and 8VC to scale Aidan, an "AI employee" that watches your browser, talks to your users, and clicks the buttons itself. Notion and Decagon are already in production. 150+ companies are on the waitlist. If you are trying to build the same thing, or hire against it, the useful question is not "is this real?" It is: how many humans on Earth have actually shipped this stack, and how do you find them before Shaun Maguire does?

The pool is roughly 23 people, not 23,000

In Refolk's index of professional profiles, only 27 people globally self-identify with "voice agent, real-time AI," and exactly one self-identifies as a "browser agent + computer use" specialist. The realistic intersection, engineers who have shipped both live browser control and sub-500ms voice in the same product, is a two-digit number.

That is the actual sourcing problem behind Sable's headline. "Interactive Intelligence," which Sable defines as the combination of real-time browser navigation, vision, voice, and video that lets an AI see, click, explain, and collaborate with a user in a shared environment, is a 2026-vintage skill. The models to do it competently did not exist twelve months ago. The people who have wired them into production customers did not exist as a category until this year.

27

Sable itself said in its release that the stack "was not possible even a few months ago." Read that literally. It is a hiring statement, not a marketing one. The cohort of engineers who have integrated Anthropic's computer-use API against a WebRTC voice pipeline in front of paying customers is measured in dozens, and most of them are already at Sable, Decagon, Hippocratic AI, or a stealth Sequoia bet.

What Sequoia actually bought for $45M

Sequoia and 8VC bought a four-person Harvard team that got Notion and Decagon into production before the seed round closed. The board seats went to Shaun Maguire (Sequoia) and Joe Lonsdale (8VC), and Maguire compared the demo to "what Stripe did for payments."

The named cofounders are Nim Ravid (CEO), Leon Chen, Linda He, and Itamar Rocha. The angel list is the tell:

• Antonio Gracias (Valor Atreides AI Fund)
• Brian Halligan and Dharmesh Shah (HubSpot cofounders)
• Scott Wu (CEO, Cognition)
• Sabrina and Evan Hahn

That is a distribution-heavy cap table. HubSpot founders and a Cognition CEO do not check into a seed for the vision transformer. They check in because Aidan replaces the human on a sales call, and every SaaS company on the waitlist has an inbound demo team they would like to shrink by 60%.

Sequoia partner Julien Bek has been arguing publicly that "services are the new software" and that the next trillion-dollar company will sell results, not tools. Sable is the purest expression of that thesis available for a $45M check. Which means the engineer profile Sable and its imitators need is not a research scientist. It is someone who can hold a Playwright loop, a Deepgram or Cartesia voice pipe, and a demo script in their head at the same time.

Why title-search will fail you here

Searching LinkedIn for "computer use agent engineer" returns consultants and hobbyists, not shippers. Refolk's single global "computer use expert" self-ID is a solo consultant in Wenatchee, Washington. That is not the pool you want.

The pool you want lives one layer down, in stack pairs that no job board indexes:

1. Playwright or Browserbase contributors who also touch LiveKit
2. Anthropic computer-use API early builders shipping in front of real users
3. pipecat contributors with a background in browser automation
4. Deepgram, Cartesia, or Rime integrators who also wrote a screen-parsing loop
5. Engineers at Agora, 100ms, or Giga who moved to an LLM company in the last 9 months

None of that surfaces on a title search. All of it surfaces on a skill-stack plus GitHub plus employer-history cross-reference, which is the exact gap Refolk closes: you describe the person in plain English ("engineers who shipped Anthropic computer-use in production and previously worked on WebRTC voice at Agora or 100ms") and get a ranked shortlist across GitHub, LinkedIn, and the open web.

The job title you want to hire for does not exist on any job board yet. Source on stack pairs, not titles.

The real farm team is voice infra, not OpenAI

The counterintuitive move is to stop raiding LLM labs and start raiding WebRTC companies. In Refolk's index, the top employers for real-time voice AI talent are voice infrastructure companies, not model shops.

Giga, Agora, 100ms, Hippocratic AI, Chambr, and isLucid are where the movable talent sits. These are the companies that have been running low-latency audio in production for years. When Anthropic shipped a computer-use API that could actually click buttons, the engineers at these companies were the only ones who already knew how to stream, buffer, and interrupt inside a 400ms budget. Their models-team peers at OpenAI did not.

If you are hiring against Sable, do not start with the ex-OpenAI list. Start with the ex-Agora and ex-100ms lists, filter for anyone who has pushed a commit to pipecat, LiveKit Agents, or Browserbase in the last six months, and work backwards.

The math on Sable's waitlist is brutal

If 150 companies are on Sable's waitlist and roughly 23 engineers globally can build a competing stack, that is one qualified builder per 6.5 companies chasing this category. Sable itself employs a meaningful fraction of that pool.

That ratio has three consequences for founders and recruiters:

• The hiring window closes at seed. Sable was less than a year old and already had Notion and Decagon in production. Waiting until Series A to start sourcing this profile means bidding against a company with a Sequoia term sheet and Scott Wu on the cap table.
• Cash does not fix a two-digit pool. At $45M raised per Sable and roughly 23 shippers globally, the effective capital-per-engineer in the category is around $1.9M and rising. The constraint is not budget. It is knowing which 23 people to call.
• Poaching is the only path. There is no bootcamp graduating this profile. There is no university program. Every hire is a poach from Decagon, Hippocratic AI, Sable itself, or a voice-infra alum who moved sideways.

1.9M

Where Sable's competitors will actually come from

The next four Sables will come out of the same three university cohorts and the same six voice-infra alumni lists. Harvard, Stanford CS, and Waterloo are producing the founder teams. Agora, 100ms, LiveKit, Deepgram, Hippocratic AI, and Decagon are producing the engineers those teams need to hire in month four.

If you are running sourcing for one of Sable's competitors, or building the internal version at Notion or Decagon (both of which now have a strong reason to insource this stack), your target list has three tiers:

1. Anyone Maguire, Lonsdale, Gracias, Halligan, Shah, or Wu has backed in the last 18 months. These investors have already seen the pattern. The engineers at their portfolio companies are pre-vetted for this exact profile.
2. The Notion and Decagon engineers who worked on the Sable integration. They know the failure modes better than Sable's own team, because they had to break Aidan in prod. In Refolk's index you can pull them by describing the integration work in plain English rather than guessing at internal titles.
3. GitHub contributors to pipecat, LiveKit Agents, Browserbase, and Anthropic's computer-use examples repo. These are the people who care enough about the problem to write code for free.

The reason a plain-English sourcing tool matters more than a Boolean builder here is that none of these people wear a title that maps to the job. A LinkedIn search for "computer use engineer" returns noise. A Refolk query like "GitHub contributors to LiveKit Agents or pipecat who currently work at a voice AI company in the US and have shipped browser automation" returns the actual list. That is the difference between a two-week sourcing sprint and a two-quarter one.

What to write in the JD (and what to delete)

Delete "5+ years of AI experience." The category is nine months old. Delete "PhD preferred." Sable's founders are undergraduates. Delete every mention of "prompt engineering." Nobody worth hiring calls it that anymore.

Write instead:

• Shipped a real-time voice loop under 500ms end-to-end latency
• Shipped a browser-control agent (Playwright, Browserbase, or Anthropic computer-use) against a live user, not a demo
• Comfortable owning both the model call and the WebRTC transport
• Willing to sit in on sales calls and watch the agent fail in front of a customer

That last bullet is the "services are the new software" filter. The engineers who win in this category are the ones who treat the sales demo as the production surface, because for Sable, it literally is. If your JD reads like a 2023 ML role, you will hire the wrong person and lose to a four-person Harvard team that already understood the assignment.

FAQ

How big is the global pool of engineers who can build a Sable competitor?

Roughly 23 people worldwide have shipped the full stack of real-time voice plus live browser control plus vision in front of paying customers. Refolk's index shows 27 self-identified real-time voice AI professionals and exactly 1 self-identified browser-use specialist, and the overlap in production shippers is a two-digit number. This is not a market you hire from at scale. It is a market you poach from one name at a time.

Where should I source before I touch OpenAI or Anthropic alumni lists?

Start with voice infrastructure companies: Agora, 100ms, Giga, LiveKit, Deepgram, and Hippocratic AI. These are the shops that have been running sub-500ms audio in production long enough to have engineers who can hold the whole pipeline in their head. Cross-reference against GitHub contributors to pipecat, LiveKit Agents, and Browserbase. The LLM labs will feel like the obvious pool, but they are downstream of the real farm team.

What query should I actually paste into a sourcing tool?

Something like: "US-based engineers who have contributed to pipecat or LiveKit Agents in the last 12 months and previously worked at Agora, 100ms, Giga, or Deepgram, and have shipped Playwright or Anthropic computer-use in production." That is the exact shape Refolk is built for, because none of those signals survive a Boolean title search on LinkedIn. You describe the person in plain English and get the ranked list across GitHub, LinkedIn, and the open web.

Is Sable's $45M round actually a signal to build a competitor, or too late?

It is a signal that the category is real and that the hiring window is closing, not that it is closed. Sequoia and 8VC just validated Interactive Intelligence as a venture-scale bet, which means the next four rounds in this space will close in the next six months. But the pool is so small that whoever locks up the second-tier voice-infra alumni first, the ex-Agora and ex-100ms engineers who have not yet moved to an LLM company, will out-hire the founders who are still writing "AI agent engineer" job descriptions.