Adept AI Closing Archive — June 28, 2024: Amazon Absorbs the AGI Lab and David Luan, the Action Transformer Outflow

MEMEX retro archive — Adept AI / Amazon official statement, June 28, 2024

Amazon official statement · June 28, 2024 "Amazon hires Adept's five co-founders plus some staff, and at the same time signs a non-exclusive license to Adept's agent technology, multimodal models, and datasets" — Adept's David Luan reports directly to Rohit Prasad, head of Amazon's AGI team.

The single thing to take away from this article is that the Adept AI soft acquisition is the second hyperscaler-style acquihire of 2024, occurring three months after Microsoft-Inflection (March 19, 2024) with the same structural shape. The moment the same pattern repeated, it stopped being "an isolated deal" and became worth recording in the MEMEX archive as "a new exit standard for the industry."

MEMEX retro archive No. 2. On June 28, 2024, Amazon announced the hire of Adept AI co-founder and CEO David Luan, plus four other co-founders and some staff, and signed a non-exclusive license with Adept. The remaining company appointed former Head of Engineering Zach Brock as the new CEO and announced a pivot to "solutions that enable agentic AI." Structurally identical to Inflection-Microsoft (March 19, 2024), it is the second hyperscaler-style acquihire of 2024. Together with the later Character.AI-Google deal (August 2, 2024), it is one of three cases MEMEX analyzes as "the 2024 soft-acquisition trio."

Industry context — why Adept is retro No. 2

Adept AI was co-founded in 2022, centered on David Luan (formerly VP of Engineering at OpenAI), as an Action Transformer (ACT) startup Adept AI's core technical concept. Where standard Transformer-based LLMs center on text generation, Action Transformer is designed to learn and predict 'actions' on software (click, type, navigate, form submit, etc.) as a sequence. The result enables browser operation, SaaS automation, and enterprise software workflow execution as an agent. The flagship model is ACT-1 (Action Transformer 1), released September 2022, followed by Fuyu-8B (multimodal, open source) in October 2023 and Fuyu-Heavy (multimodal, enhanced UI understanding) in January 2024. . The original co-founders included not only David Luan (CEO) but Ashish Vaswani and Niki Parmar — co-authors of the Transformer paper "Attention Is All You Need." Both left early to co-found another startup (source: TechCrunch, June 28, 2024 report), and Adept continued with the remaining five co-founders (David Luan, Augustus Odena, Maxwell Nye, Erich Elsen, Kelsey Szot). Total funding reached around $415 million (per reporting), with major VCs like General Catalyst and Greylock participating, focused on the enterprise B2B segment.

On June 28, 2024, Amazon officially announced the hire of five co-founders, including David Luan, plus some staff, into Amazon's AGI lab. At the same time, it signed a non-exclusive license for Adept's agent technology, multimodal models, and some datasets. The remaining company appointed former Head of Engineering Zach Brock as the new CEO and announced a pivot to "solutions that enable agentic AI." This structure is fully isomorphic with Inflection-Microsoft (March 19, 2024) three months earlier — (a) a hyperscaler absorbs the principal technical team via license + hire, (b) the remaining company continues as a separate business with a B2B / enterprise pivot, (c) the form sidesteps M&A scrutiny under antitrust law.

Factual timeline (primary-source-based)

  • Early 2022: Adept AI co-founded (David Luan, Ashish Vaswani, Niki Parmar, Augustus Odena, Maxwell Nye, Erich Elsen, Kelsey Szot). Vaswani and Parmar left during 2022 to co-found another startup (source: TechCrunch, June 28, 2024 report)
  • September 14, 2022: ACT-1 (Action Transformer 1) released — the first serious model centered on predicting actions on software (source: Adept official blog, "ACT-1: Transformer for Actions")
  • March 2023: Series B raise of $350 million (led by General Catalyst / Spark Capital), valuation around $1 billion (per reporting)
  • October 2023: Fuyu-8B released, a multimodal model specialized for UI understanding, open-sourced (source: Adept official blog, "Adept Fuyu-8B")
  • January 24, 2024: Fuyu-Heavy released, claimed by Adept to be the multimodal model with performance second only to GPT-4V / Gemini Ultra (source: Adept official blog, "Adept Fuyu-Heavy")
  • June 28, 2024: David Luan, four other co-founders, and some staff are hired into Amazon's AGI lab. Adept and Amazon simultaneously sign a non-exclusive license (covering agent technology, multimodal models, and datasets). Zach Brock (former Head of Engineering) becomes Adept's new CEO. There is no Amazon official announcement page on the day; Rohit Prasad's (head of Amazon's AGI team) internal memo, obtained by GeekWire, was reported the same day by GeekWire, CNBC, and TechCrunch, propagating the news through the industry (source: GeekWire, CNBC, TechCrunch, all June 28, 2024)
  • June–August 2024: Remaining Adept advances a pivot to enterprise B2B under the slogan "solutions that enable agentic AI." Semafor reports that Amazon's license fee was applied to repaying Adept investors (the remaining company "returned capital and continued as a separate business")
  • December 2024: Amazon officially announces a new AGI-related lab led by Adept co-founder David Luan (source: TechCrunch, December 9, 2024)
  • February 2026: Amazon AGI lab head Rohit Prasad reported to be leaving; David Luan's position inside Amazon continues (source: CNBC, February 24, 2026)

Why it disappeared (= exited the frontier) — structural analysis

Adept AI has not disappeared as a corporate entity, but it has exited frontier agent model development. MEMEX reads three structural reasons from primary sources:

(1) Mismatch between funding scale and compute unit cost. Adept raised around $415 million in total, but Action Transformer–style agent model development requires continuous multimodal training, collection of real-software-operation data, and reinforcement learning, with compute costs higher than for simple LLM training. In the same period, Anthropic's Claude (with built-in computer use) and OpenAI (Operator-style agent products) entered the market with hyperscaler backing, and it is likely Adept judged $415 million insufficient to sustain direct competition.

(2) Difficulty of product-market fit in the enterprise B2B segment. Adept centered on enterprise software automation from the start, but the same enterprise agent market saw Microsoft Power Automate, Salesforce Agentforce, ServiceNow Now Assist, and various RPA vendors (UiPath, Automation Anywhere) layer AI features and enter. It is structurally difficult for a pure AI agent startup to sustain differentiation in direct competition with established SaaS vendors.

(3) Alignment of interests with Amazon's AGI strategy. Amazon serves Anthropic Claude via AWS Bedrock, while also continuing to train its own frontier models in the Titan / Olympus line. Amazon, however, has been structurally behind in consumer products versus OpenAI's ChatGPT and Google Gemini, and incorporating David Luan, the co-founders, and the Action Transformer technology stack carries meaning as a way to rapidly strengthen Amazon's AGI team's agent capability — so MEMEX reads (this is the editor's inference; Amazon has not officially detailed its AGI strategy).

What the industry was signaling — establishing the "same-pattern" second case of 2024

The significance of the Adept-Amazon case is that two structurally isomorphic deals occurred within the same year, three months apart — Inflection-Microsoft and then Adept-Amazon. The second occurrence settled the observation hypothesis: "this is not isolated; it is the industry's new exit pattern." MEMEX observes five common structures:

  • A hyperscaler absorbs co-founder / CEO + principal technical team via license + hire
  • The remaining company continues as a separate business with a B2B / enterprise pivot
  • The form sidesteps M&A scrutiny under antitrust law (licensing + hiring rather than equity acquisition)
  • The license fee is paid to the remaining company, providing a partial path to capital recovery for investors
  • It effectively exits frontier model competition, but leaves room for the remaining company to persist in the enterprise market

With a third case in the same year (Character.AI-Google, August 2, 2024) reproducing the structure, MEMEX records 2024 in the archive as "the watershed year in which independent startups competing at the frontier converged into a hyperscaler oligopoly." For details, see The 2024 soft-acquisition trio — the watershed in which Microsoft, Amazon, and Google ran the same format three times within six months.

Editorial reading — the Adept case as a MEMEX industry axis

(1) The survivability of Action Transformer technology. Adept's ACT-1 / ACT-2 were unique models specialized for software workflow automation, but equivalent functionality was standardized in 2024–2025 as Anthropic's Claude Computer Use, shrinking the differentiation value of ACT technology in isolation. This is an early case in the industry pattern MEMEX observes: "features of specialized AI startups are absorbed as standard features of frontier general models."

(2) Bifurcation of the enterprise agent market. After the Adept-Amazon soft acquisition, the enterprise agent market split into two poles: (a) direct provision by hyperscalers (Amazon Bedrock Agents, Microsoft Copilot, Google Vertex AI Agents), and (b) specialized vertical agents (legal, sales, marketing, etc.). The space for an independent startup as a pure agent runtime is structurally narrowing, and vertical-specialized players (e.g., Intercom's company-wide Claude Code rollout, Anthropic Cowork role-based demos) are becoming central.

(3) The meaning of taking up Adept in the retro series lies in self-verifying the observation hypothesis of "establishing the same pattern" at the second occurrence. A single case (Inflection alone) can also be read as "an isolated deal," but the moment a second isomorphic case appears in the same year, the isolation hypothesis is rejected. This is the core of MEMEX editorial posture — "depth that will be cited months to years later" — completed not by the Adept retro alone, but reinforced by its position within the trio.

Related resources

Sources

Primary-source equivalent (Adept official blog and party-of-record materials)

Secondary sources (treated as industry archive)

Editor's note: Amazon has not published an official announcement page for this on aboutamazon.com or the Amazon Newsroom. The path of transmission to the industry was Rohit Prasad's (head of Amazon's AGI team) internal memo, obtained and published by GeekWire, which has been treated as primary-source equivalent by the industry. In addition, the dollar amount of the Amazon-Adept license was not officially disclosed by either party. Following retro editorial rules, the absence of an official announcement is made transparent, and industry primary-source-equivalent sources are classified separately as "primary-source equivalent."