Amazon's July AGI Cut Was Undisclosed. The Real Pool Is 73.

Amazon confirmed on July 22, 2026 that it was cutting roles inside its AGI organization and declined to say how many. That silence is the whole story: no WARN filing, no headcount, no press-friendly round number, and therefore no external recruiter queue. The pool is small, named, and mostly still updating profiles.

What Amazon actually cut in July 2026

Amazon eliminated roles across two named AGI sub-orgs and shut down its 18-month-old AGI SF Lab, without publishing a headcount. The unit was already bleeding leadership before the cut.

The confirmed facts, in the order they matter to a sourcer:

• July 22, 2026: Amazon confirmed layoffs inside the AGI organization but declined to disclose numbers or which parts were affected (CNBC).
• Sub-orgs named as hit: AGI Data Services (VP Adeeb Shanaa) and AGI Information (VP Vishal Sharma), per Business Today.
• AGI SF Lab closed outright. The lab was set up in December 2024, grew to roughly 80 people at peak per The Information, and shipped Nova Act, the browser-agent model still available on AWS.
• Leadership departures preceded the cut. Rohit Prasad was replaced by Peter DeSantis at the top of the AGI org in December 2025. David Luan, who ran the AGI Lab after Amazon acquihired Adept, left in February 2026.
• Adept diaspora already leaking. More than a dozen ex-Adept hires left before July, per GeekWire.

The contrast with January matters. Amazon's January 2026 wave cut over 16,000 roles and every recruiting team in the country had a spreadsheet on it within a week. July got no such spreadsheet, because Amazon never gave anyone a number to put in cell A1.

16,000

Why the "quiet cut" framing hides a mechanism, not a mystery

Amazon didn't disclose a number because the July action wasn't one action. It was a lab closure plus targeted cuts inside two VP orgs, which is legally and narratively easy to under-report.

That has three sourcing consequences:

1. No WARN trigger for the SF portion at reportable scale. The AGI SF Lab peaked around 80, and more than a dozen ex-Adept hires had already exited before July. The remaining SF headcount at closure was almost certainly below California's WARN threshold at any single address once you subtract remote and Seattle-attached staff.
2. The Seattle cuts were absorbed into the January narrative. Puget Sound recruiters are still working the January list; the July delta rounds to zero against 16,000.
3. The pool is only identifiable by org-graph triangulation. Nova paper co-authorship, Nova Act model card credits, AGI-SF arXiv publications, and pre-2024 Adept GitHub history are the only reliable filters. Title strings alone under-count badly, because Amazon's internal titles for this org drift ("Applied Scientist," "Research Scientist," "Software Engineer, Amazon AGI," "Senior ML Engineer, AGI Data Scaling," "Head of Product, Agentic AI, Amazon Nova").

This is the exact gap Refolk closes for ex-Amazon AI engineers hiring: you describe the person in plain English ("applied scientists who co-authored Nova Act or worked in Amazon AGI SF, US-based, open to move") and get a ranked shortlist that crosses GitHub, LinkedIn, and arXiv without you writing a boolean.

The actual pool: 73 profiles, 16 still at Amazon

In Refolk's index of professional profiles, 73 US-based Applied Scientists and Research Scientists carry "Amazon AGI" in their headline. Sixteen still list Amazon as their current employer, which is the cohort most likely to include July-cut engineers who haven't updated yet.

A narrower slice on Nova foundation-model work returns 15 profiles, 11 of them still at Amazon, with titles like "Software Engineer, Amazon AGI," "Senior ML Engineer, AGI Data Scaling," and "Head of Product, Agentic AI, Amazon Nova." Regional distribution skews Seattle, not SF: 5 of the 15 Nova-tagged profiles are in Seattle, 2 in the Bay Area.

Two things jump out. First, the reachable pool is measured in dozens, not hundreds. Second, the demand side is measured in hundreds. That mismatch is the entire recruiting problem.

The reachable pool is measured in dozens. The demand side is measured in hundreds. That is the entire recruiting problem.

The 1:8 displacement-to-demand ratio

Meta Superintelligence Labs alone has roughly 8 times the Amazon AGI displaced pool in Refolk's index, before you add Anthropic, xAI, or any well-funded seed lab. Amazon AGI SF recruiting in August 2026 is not a package fight, it is a speed fight.

The math:

• 73 ex/current "Amazon AGI" scientists in the US.
• 605 US profiles at Meta Superintelligence Labs, with top-cited titles of Research Engineer and Software Engineer, meaning they are hiring for the exact seat these people just left.
• ~1:8 ratio, before Anthropic, which has been running a parallel absorption on the RLHF and safety side, and before Patronus AI and Luma, both of which already appear as landing spots for ex-Amazon-AGI scientists in the Refolk sample.

Meta will not feel the absorption. You will.

How to identify the July cohort without a WARN filing

You identify the July cohort by triangulating three public artifacts: Nova Act attribution, Adept-era GitHub history, and AGI-SF publication co-authorship. Titles alone will miss half of them.

The five filters that actually work:

1. Nova Act model card and blog credits. Nova Act still ships on AWS. Anyone named on its research or engineering track was inside AGI SF.
2. Ex-Adept commit history. Amazon acquihired roughly 66% of Adept's staff in June 2024 after Adept raised about $400M. The GitHub org history and arXiv co-authorship from 2022 to mid-2024 is the cleanest Adept-alum filter.
3. AGI-SF arXiv publications, Dec 2024 to July 2026. The lab was barely 18 months old, so the paper set is small and readable in an afternoon.
4. Reports-into signal for the two named VPs. Adeeb Shanaa (AGI Data Services) and Vishal Sharma (AGI Information) had public org shapes visible via LinkedIn second-degree graphs.
5. David Luan's follow-graph. Luan left in February 2026. The engineers who followed him out over the next six months are a leading indicator, not a lagging one, for who is open to a call now.

Sourcing AGI researchers this way, by hand, takes a senior sourcer roughly a full week per 20 profiles. Refolk collapses that to a plain-English query and a ranked list, because the index already carries the Nova Act, Adept, and AGI-SF publication signals as features on each profile.

Named entities worth putting on your board today

The people to know, sorted by sourcing utility rather than seniority:

• David Luan. Former Adept CEO, ran Amazon's AGI Lab, exited Feb 2026. Anchor node for the ex-Adept diaspora.
• Pieter Abbeel. UC Berkeley professor, joined AGI SF Lab as a researcher. His co-author graph is a shortlist generator.
• Rohit Prasad. Former head of the AGI org, replaced by Peter DeSantis in Dec 2025. His org chart from 2024 is the Amazon-Nova team engineers map.
• Adeeb Shanaa (VP, AGI Data Services) and Vishal Sharma (VP, AGI Information). Their direct reports are the July impact zone.
• Nova Act core team. Named on the model card, still shipping on AWS, all verifiable.

Destinations already showing up in the Refolk sample as landing spots for ex-Amazon-AGI scientists: Meta Superintelligence Labs, Anthropic, Patronus AI, and Luma. If you are recruiting for one of these, you are already competing with the others; if you are a Series B lab that is not on this list yet, you have about four weeks before every remaining candidate is in a Meta loop.

The four-week window, priced

The window closes roughly four weeks from the July 22 announcement, which puts the practical deadline in late August 2026. Every week past that raises your per-hire cost against Meta Superintelligence and Anthropic, not because packages inflate but because the top of the pool disappears into signed offers.

Concretely:

• Week 1 to 2 (July 22 to Aug 5): Passive candidates test the market. Warm intros through the Adept diaspora convert cheapest here.
• Week 3 to 4 (Aug 6 to Aug 19): Meta and Anthropic first-round loops start closing. Package escalation begins.
• Week 5+ (Aug 20 onward): The 16 "still at Amazon" profiles in the Refolk slice update to new employers. Your outreach hit rate on that cohort drops sharply.

If you are running Amazon AGI SF recruiting or hunting Amazon Nova team engineers, the practical move this week is to pull the 73-person list, filter to the 16 who haven't updated, and send first-touch messages that reference a specific Nova Act artifact or arXiv paper. Generic "saw you were at Amazon" notes will get ignored, because these candidates are getting 40 of them a day.

FAQ

How many people did Amazon actually lay off from the AGI team in July 2026?

Amazon did not disclose the number and explicitly declined to when asked on July 22, 2026. Public reporting confirms roles were cut inside AGI Data Services and AGI Information, and that the AGI SF Lab was closed. The SF lab peaked at about 80 people per The Information, and more than a dozen ex-Adept hires had already left before July, so the SF-attributable cut was a fraction of 80. Refolk's index shows 73 US profiles tagged "Amazon AGI" in scientist titles, with 16 still listing Amazon as current employer, which is the practical size of the reachable pool.

Where is the ex-Amazon AGI talent concentrated geographically?

Seattle and Bellevue outweigh the San Francisco Bay Area for this cohort by roughly 2 to 1 in Refolk's Nova-tagged slice: 5 of 15 Nova-tagged US profiles are in Seattle versus 2 in the Bay Area. The AGI SF Lab was the visible SF footprint at about 80 peak, but the broader AGI org sits in Puget Sound. Sourcing playbooks that assume "AGI equals SF" will miss the majority.

Who is competing for these engineers right now?

Meta Superintelligence Labs is the largest visible absorber, with roughly 605 US profiles in Refolk's index, top titles Research Engineer and Software Engineer. Anthropic is running parallel absorption, particularly on safety and RLHF-adjacent roles. Patronus AI and Luma already appear as landing spots for ex-Amazon-AGI scientists in the Refolk sample. That is at least four active buyers against a displaced pool measured in dozens.

What is the fastest way to build a shortlist of the July cohort?

Triangulate on Nova Act attribution, ex-Adept GitHub history, and AGI-SF arXiv co-authorship, then cross-reference against the two named VP orgs (Adeeb Shanaa and Vishal Sharma). Doing this by hand takes a senior sourcer about a week per 20 profiles. A plain-English query in Refolk returns the same shortlist in minutes because Nova Act, Adept, and AGI-SF publications are already indexed as signals on each profile.