Adept

Adept is an American AI research lab founded in 2022 by a cohort of former OpenAI and Google researchers, including three of the eight co-authors of the original transformer paper, that built agentic-AI models for software workflow automation before its senior leadership was acqui-hired by Amazon in June 2024. The company persists as a smaller independent entity under new chief executive Zach Brock, focused on enterprise agentic AI solutions, while the founding research talent and the company's core foundation-model technology were licensed to Amazon to seed the new Amazon AGI organisation.

At a glance

• Founded: April 2022 by David Luan, Ashish Vaswani, Niki Parmar, Augustus Odena, Maxwell Nye, Erich Elsen, Kelsey Szot, Anmol Gulati, Kelvin Guu, and Fred Bertsch.
• Status: Private. Operating as a substantially smaller entity after the June 2024 Amazon acqui-hire.
• Funding (pre-acqui-hire): Approximately $415 million raised across seed, Series A, and Series B rounds. Series B closed at a $1 billion post-money valuation in March 2023.
• Notable investors: Greylock, Spark Capital, Addition, Nvidia, Atlassian, Workday, General Catalyst.
• Founding chief executive: David Luan (April 2022 to June 2024). Departed for Amazon in the acqui-hire and reports to Rohit Prasad as part of the Amazon AGI organisation.
• Current chief executive: Zach Brock, formerly Adept's head of engineering, promoted to CEO during the Amazon transaction.
• Open weights: Persimmon-8B (September 2023) and Fuyu-8B (October 2023). Both released under permissive licenses.
• Flagship products: ACT-1 (March 2022 demonstration), ACT-2 (internal development), Persimmon-8B, Fuyu-8B (multimodal). No public general-availability product release reached market before the acqui-hire.

Origins

Adept was founded in April 2022 in San Francisco around the thesis that a "general intelligence for getting things done on a computer" was the next major capability frontier, and that the most economically valuable application of large language models would be in automating the keystroke-and-mouse-click workflows that knowledge workers perform inside desktop and web software. The founding team was unusually heavy on transformer-paper credentials: Ashish Vaswani (lead author of "Attention Is All You Need"), Niki Parmar (co-author), and Anmol Gulati (later co-author on Conformer and related architectures) were three of the founders. The remaining co-founders included David Luan (former vice president of engineering at OpenAI and former Google director), Augustus Odena (former Google Brain research scientist), Maxwell Nye (former OpenAI researcher), Erich Elsen (former DeepMind and Google Brain researcher), Kelsey Szot (former Google Brain), Kelvin Guu (former Google), and Fred Bertsch (former Google).

The founding cohort's specific arc is significant. Vaswani, Parmar, and Gulati left Google Brain in 2021 with the explicit thesis that the next research-and-product company built around the transformer architecture should be founded by the people who designed it. Luan brought the operator and product credentials, and the broader founding team rounded out the research-and-engineering staff. The launch event in April 2022 paired the founding announcement with a Series A round, an unusual sequencing for an AI startup at the time and a signal of how compressed the talent-to-capital lag was even in the pre-ChatGPT period.

Adept's launch demo product was ACT-1, the Action Transformer, released as a public demonstration video in March 2022. The demo showed an AI agent operating a web browser to complete real software tasks (filling forms, navigating multi-page workflows, executing search-and-action sequences) under natural-language instruction. The technical premise behind ACT-1 was that the right approach to agentic AI was a foundation model trained directly on action-and-observation trajectories rather than a language-model-with-tool-use scaffolding.

Ashish Vaswani and Niki Parmar departed Adept in 2022, before the company's first major release cycle, to pursue a separate research-and-foundation-model thesis. The two founded Essential AI in 2023. Their departure was the first significant founding-team change at Adept and is the diaspora-map data point that the company's structural cohesion was never tested at full strength.

Mission and strategy

Adept's stated mission was to build "general intelligence for getting things done on a computer," operationalised through agentic models that could observe screens, take actions, and complete multi-step software workflows under natural-language instruction. The strategic premise was that the foundation-model-for-action approach (training directly on action-and-observation data) would produce more reliable agents than the language-model-plus-scaffolding approach the rest of the field was pursuing.

The pivot toward enterprise productivity in 2023 reflected a difficulty translating the original consumer-facing-agent demo into a shippable product. ACT-1's launch demonstration was visually impressive but the underlying reliability at production scale proved harder to achieve than the demo suggested, and the consumer-agent market was crowded with both incumbent products (the various RPA vendors, the early consumer assistants) and adjacent foundation-model-plus-scaffolding agents from competitors. The enterprise pivot positioned Adept as an integration-and-deployment partner for large companies rather than as a consumer-product company.

The transition was incomplete. Adept never released a generally-available product under its original brand before the Amazon transaction, and the enterprise traction was sufficient to fundraise but insufficient to generate the kind of revenue trajectory that would have justified a stand-alone Series C round at the company's last valuation.

Models and products

• ACT-1 (March 2022). The Action Transformer demo, the company's launch event and most-visible public artefact. Public demo only; no generally-available product release.
• ACT-2 (in development, never publicly released). Successor to ACT-1, intended to address the reliability and scale limitations of the original demo. Capability details were limited to selected enterprise-partner demos and internal evaluations.
• Persimmon-8B (September 2023). Open-weights 8-billion-parameter language model, released to position Adept inside the open-weights research conversation and to provide a baseline for the company's foundation-model work. Apache 2.0 licence.
• Fuyu-8B (October 2023). Open-weights multimodal model with native image-input capability, distinguished by its architectural simplicity (single decoder-only transformer with image patches as input tokens). The Fuyu architecture was an early demonstration of the patch-as-token approach that subsequent multimodal models have built on.

No enterprise-AI software product reached general availability under the Adept brand before the Amazon transaction.

Benchmarks and standing

Adept's published benchmark numbers were primarily around the Fuyu and Persimmon releases. Fuyu-8B reported competitive performance on standard multimodal benchmarks for its parameter scale at release, though without claiming frontier leadership. Persimmon-8B reported standard general-language-model benchmarks at the 8-billion-parameter scale; the release was understood as a research-and-recruiting artefact rather than as a competitive frontier release.

The benchmark that would have mattered most for Adept's strategic position (reliable end-to-end software workflow automation on production enterprise environments) does not have a comparable standard benchmark, and the company's internal evaluation data has not been publicly disclosed.

Leadership

• David Luan (Co-founder, former CEO, April 2022 to June 2024): Former vice president of engineering at OpenAI and former Google director. Currently a vice president at Amazon reporting to Rohit Prasad, head of the Amazon AGI organisation.
• Ashish Vaswani (Co-founder, departed 2022): Lead author of "Attention Is All You Need." Co-founded Essential AI in 2023.
• Niki Parmar (Co-founder, departed 2022): Transformer-paper co-author. Co-founded Essential AI in 2023; subsequently joined Anthropic in December 2024.
• Augustus Odena, Maxwell Nye, Erich Elsen, Kelsey Szot (Co-founders): Departed for Amazon AGI in June 2024 alongside David Luan.
• Zach Brock (current CEO): Former head of engineering, promoted during the Amazon transaction to lead the residual entity.

Funding and backers

Adept's pre-acqui-hire funding totalled approximately $415 million across three rounds. The seed round in early 2022 was led by Addition. The Series A in March 2022, announced concurrently with the company's public launch, was led by Greylock and Addition with participation from Nvidia, Atlassian, and others; the round size was approximately $65 million. The Series B in March 2023 was approximately $350 million at a $1 billion post-money valuation, led by General Catalyst and Spark Capital with participation from Nvidia, Atlassian, Workday, and Microsoft. The Series B was one of the largest 2023 enterprise-AI rounds and priced Adept among the well-capitalised insurgents at the time.

The Amazon transaction in June 2024 was not a traditional acquisition. Amazon hired the senior founding team and licensed Adept's AI technology (specifically the agentic-and-multimodal-foundation-model work) into the Amazon AGI organisation, while Adept itself remained an independent company with the residual staff, IP not included in the Amazon licence, and a continuing-operator commitment. The financial terms of the licence were not publicly disclosed. Industry coverage described the transaction as effectively a partial acqui-hire that gave Adept's investors a partial-liquidity event while preserving the residual entity for continued operation.

Industry position

Adept's industry position before the Amazon transaction was as a foundation-model-grade insurgent in the agentic-AI category, alongside competitors that pursued similar mandates from different starting points (Anthropic's tool-use focus, OpenAI's GPTs and operator framings, the various enterprise-RPA-plus-LLM integrations). After the transaction, the residual Adept entity is best understood as a smaller enterprise-AI integration company rather than as a foundation-model competitor. The brand-level positioning has been substantially reduced in industry coverage since mid-2024, with Adept appearing primarily in retrospective coverage of the Amazon deal rather than in active competitive analysis.

The Amazon AGI organisation that the senior team joined is now itself the more relevant strategic actor in the agentic-AI category, with the Adept-acquired talent as a significant portion of its founding research leadership.

Competitive landscape

• Amazon AGI (US, post-acqui-hire successor): The organisation that absorbed Adept's founding leadership and licensed its technology. Currently the more important strategic actor on the Adept-research line.
• Anthropic (US, tool-use and Claude Operator agent): Direct competitor on enterprise agentic AI, with stronger product traction and a clearer commercial-deployment story.
• OpenAI (US, GPTs and Operator): Closed-frontier-model competitor with agent-product offerings (Operator in particular) competing in the same agentic-AI category.
• Sierra (US, enterprise customer-AI agents): Bret Taylor's enterprise-AI startup, a different positioning (customer-facing rather than internal automation) but a competitor for the broader enterprise-AI-agents budget.
• Essential AI (US, foundation-model insurgent): Founded by Adept's departed co-founders Vaswani and Parmar; pursuing a different research-and-foundation-model thesis from the Adept enterprise-agent direction.
• Various enterprise-RPA-plus-LLM vendors (UiPath, Automation Anywhere, Microsoft Power Automate with Copilot integration): The incumbent agentic-AI-for-enterprise category that Adept was competing into.

Outlook

Open questions and watchable signals over the next 6 to 18 months:

• Adept's continued operation. Whether the residual entity under Zach Brock builds a sustainable independent business in enterprise agentic AI or whether it gradually winds down is an open question. The company has stated it is not closing operations, but the public-facing brand-and-product activity has been muted since the Amazon transaction.
• The Amazon AGI trajectory. The team that left Adept now anchors the senior research leadership of Amazon's AGI organisation. The agentic-AI capability stack David Luan and the founding team built at Adept is now Amazon's, and Amazon's trajectory in the frontier-AI category over 2026 and 2027 will be the more important measurement of what the Adept research line eventually produces.
• Whether the Adept brand survives. Smaller integration-and-services companies that emerge from an acqui-hire transition can persist for years or can quietly wind down within twelve to eighteen months. The 2026 to 2027 window will resolve which path Adept takes.
• The licensed technology's deployment surface. Amazon licensed Adept's foundation-model and agent technology. Whether that technology surfaces in Amazon's product line (Bedrock, the various Amazon-managed-AI services, Alexa-related agentic features, enterprise productivity tools) as a recognisable continuation of the Adept research line, or whether it gets fully absorbed into the Amazon AGI organisation's broader research pipeline without a visible Adept-line product trace, is the most informative signal on what the original founding research was worth.

Sources

• Amazon hires founders away from AI startup Adept. TechCrunch coverage of the June 2024 acqui-hire with named individuals, financial-terms context, and Amazon AGI reporting structure.
• Adept's ACT-1 launch announcement. The original March 2022 demonstration video and accompanying announcement.
• Persimmon-8B model card. The September 2023 open-weights language model.
• Fuyu-8B model card. The October 2023 open-weights multimodal model.
• Companion essay: The diaspora map for the broader cohort-fork patterns that the Adept founding and acqui-hire arcs illustrate at the agentic-AI tier. The Vaswani-and-Parmar departure to Essential AI and the later Luan-cohort departure to Amazon AGI are both captured in the people-layer transitions the essay analyses.
• Companion profiles: David Luan, Ashish Vaswani, Niki Parmar, and Amazon AGI for the people-and-destination context.