AI Wiki

Explore AI Wiki

Citation and evidence

AGI, Inc.

19 min full readUpdated 43 references

This article's verification

Report a problem with this article

More

Use this article

Raw MarkdownExplore connections

Improve this page

Suggest editRevision history

Topics

AI AgentsAI CompaniesAI Research

Cite this article

AGI, Inc. (also written AGI Inc.; website theagi.company) is a San Francisco artificial intelligence company that builds small "action" models and an assistant app that operate a phone or computer through its screen, reading what is displayed and then tapping, typing and moving between apps on a user's behalf.[1][2] The company says it was founded in 2025 in San Francisco by Div Garg, Steve Frey and Naman Garg.[1] Div Garg previously co-founded the web-agent startup MultiOn, and he has described AGI, Inc. as a lab spun out of that company.[3]

The company's products include AGI-0, a family of models it says ranges from 2 billion to more than 20 billion parameters; the AGI assistant app for Android, with an iOS preview; AGI-RUN, an inference engine for phones and other edge hardware; and a Model Context Protocol (MCP) server that lets outside AI agents control a phone.[1][2][4] It also published REAL, a benchmark built from replicas of popular websites, which appeared in the NeurIPS 2025 Datasets and Benchmarks Track.[5][6] In November 2025 Forbes reported that AGI, Inc. was in talks to raise $50 million at a $500 million valuation, after previously raising close to $10 million from investors including Menlo Ventures.[7] The company lists Qualcomm, Lenovo and Visa as partners.[1]

Overview

FieldDetail
NameAGI, Inc. (also styled AGI Inc.)[1]
Founded2025, San Francisco (company statement)[1]
FoundersDiv Garg (CEO), Steve Frey (President), Naman Garg (CTO)[1]
HeadquartersSan Francisco, California[1]
FocusOn-device "action" AI that operates phones and computers through their user interfaces[1][2]
Main productsAGI-0 models, AGI app, AGI-RUN inference engine, AGI MCP, REAL benchmark and AGI SDK[1][4][8]
Funding reported$8 million pre-seed/seed round in June 2025 (Entrepreneur); close to $10 million raised as of November 2025 (Forbes)[3][7]
ValuationNot disclosed; Forbes reported talks at a $500 million valuation in November 2025[7]
Investors namedMenlo Ventures, Anthropic, Point72 Ventures and angel investors[1][9][10]
Websitetheagi.company[11]

Expanded article table

The name is shared with an unrelated company. AGI Inc, a Cayman Islands company that refers to itself as Agibank and operates in Brazil, listed Class A shares on the New York Stock Exchange (symbol AGBK) through an initial public offering in February 2026.[12] Its financial figures do not apply to the AI company described here.

History

Background at MultiOn

Div Garg co-founded MultiOn, which Forbes described as having built "one of the first web browser agents."[7] In August 2024 MultiOn and Stanford University researchers published "Agent Q," a framework for self-improving web agents that combines guided Monte Carlo Tree Search, self-critique and off-policy Direct Preference Optimization. The paper lists the company affiliation as "The AGI Company (MultiOn)," and its authors include Naman Garg and Divyansh (Div) Garg.[13]

On December 20, 2024, Garg wrote on X that he would be "moving on from @MultiOn_AI to explore new things" and would remain an advisor to the team.[14] In an April 2026 interview he said he had run MultiOn for about two years and added: "Recently, I spun a new lab out of that company, and that became AGI Inc."[3] The multion.ai web address now redirects to AGI, Inc.'s website.[15]

2025: benchmark research and first funding

One of AGI, Inc.'s first public projects was REAL (also called REAL Bench), released in April 2025.[8] The accompanying paper, with Divyansh Garg as first author and authors from AGI, Inc., Stanford University, the University of Oxford and other institutions, describes deterministic replicas of 11 widely used websites and 112 tasks that mix information retrieval with actions that change a site's state.[5] The company also released an evaluation harness as the open-source AGI SDK Python package.[8][16] In 2025 Div Garg and Naman Garg taught "Building AI Browser Agents," a DeepLearning.AI short course built in partnership with AGI, Inc. that covers web agents and the Agent Q method.[17]

In a May 2025 essay, Garg described the company as "part foundational research lab, part ambitious startup," with a stated goal of making general-purpose AI "safe, reliable, and universally accessible."[18] Entrepreneur later reported that the company raised its first funding, an $8 million pre-seed/seed round, in June 2025.[3]

In August 2025 the company announced a partnership with Visa on "agentic commerce," meaning purchases and bookings carried out by agents for a user. Its announcement did not describe financial terms or specific Visa products.[19] In October 2025 it published two benchmark claims: 97.4% task success on AndroidWorld for its AGI-0 mobile agent and 76.26% on OSWorld for a computer-use agent it called OSAgent (see the benchmark table below).[20][21]

Fundraising talks and the Android app

On November 19, 2025, Forbes reported that AGI, Inc. was in talks to raise $50 million at a $500 million valuation, citing people familiar with the matter. Forbes said the company had previously raised close to $10 million, and its subheadline described AGI, Inc. as a "neolab." According to Forbes's sources, the company had told investors it had begun partnership discussions with device makers, and Samsung, Lenovo and Xiaomi were among the companies exploring its technology.[7]

In December 2025 The New York Times published a story by Cade Metz on startups that build copies of popular websites to train AI agents. It named AGI and Plato among the startups building such replicas so that AI systems could learn to navigate the internet and complete tasks like booking flights. The story also reported that United Airlines' lawyers had sent a takedown notice over AGI's replica of United.com, after which Garg renamed the site "Fly Unified" and removed the United logo.[22][23] In a January 2026 essay titled "Apps Are Dead," Garg argued for a "universal agent" that can operate any app or website by "looking at pixels and clicking and typing."[24]

At CES 2026 the company demonstrated its Android app, then in private beta behind a waitlist. TechCrunch reported that the app could take requests such as calling an Uber, changing the wallpaper or adding items to a DoorDash order. It said the company trained its own models with reinforcement learning, simulated Android apps in-house, and also used data from human annotators. The app used Android accessibility features, with the user's permission, to navigate and act. The models then ran on cloud GPUs, and the company told TechCrunch it expected to run the whole model on the device within about six months. TechCrunch also reported $10 million in backing from Menlo Ventures, Anthropic and other investors, and a waitlist of 30,000 people.[9]

2026: device partnerships

On March 3, 2026, AGI, Inc. announced a collaboration with Qualcomm Technologies to optimize its vision-language-action models and execution stack for Snapdragon processors, with full on-device processing as the stated goal.[25] The announcement quoted Vinesh Sukumar, a Qualcomm vice president, and Enterprise AI World reported that the platform was "available now to select partners."[25][26] At Mobile World Congress (MWC) 2026 in Barcelona, where the company exhibited in Hall 7, it described itself as "a neolab developing on-device superintelligence for mobile and computer-use."[10]

Also at MWC, AGI, Inc. showed a limited proof-of-concept inside Lenovo Qira, in which an agent booked a restaurant, ordered a ride and navigated checkout flows on Android using each app's own interface. The company called it "a controlled proof-of-concept demonstration and not a product announcement," and said full on-device execution was "still ahead."[27] In a conversation with Garg recorded at MWC and published by Counterpoint Research in April 2026, the firm summarized the company's work as optimizing models of under 1 billion to 8 billion parameters for Snapdragon chips and building custom models for device makers such as Lenovo.[28]

Tom's Guide tested the app, still in private beta, in April 2026. It reported that the app used a vision-language model to read the screen and accessibility features to act, and that booking an Uber took about 15 seconds from prompt to the final approval step. The reviewer noted that the agent still ran on the company's cloud GPUs.[29] In April 2026 Entrepreneur reported that the company was raising another round, and Garg said it was preparing its first iOS product. He also gave his own figures for the business: about 500,000 waitlist sign-ups, at least $1 million in revenue, and an expectation of revenue "in the $20 million range" for 2026.[3]

Through mid-2026 the company published work on running models on the device itself. Its May 2026 post described a compiler for processing-near-memory hardware developed by Adam Awale with the MPS Lab at Arizona State University, and its July 2026 post introduced AGI-RUN.[30][31] The company said in July 2026 that an accessibility professional had tested the app against WCAG 2.2 A and AA. It also said it had shipped a verbose narration mode for blind and low-vision users, and that it would host a booth at the Abilities Expo in Houston.[32] In August 2026 it announced that Android co-founder Rich Miner had joined as an advisor and released its MCP server in developer beta.[33][4] In September 2026 it updated its AndroidWorld leaderboard entry to 100%.[34]

Founders and leadership

PersonRoleBackground
Div GargCo-founder and CEO[1]Co-founded MultiOn.[3] Forbes, citing LinkedIn, described him in November 2025 as on leave from a computer science PhD at Stanford;[7] in April 2026 he said he had dropped out of the program.[3] His website says he created and taught Stanford's CS 25: Transformers United seminar and lists research on imitation learning (IQ-Learn) and offline reinforcement learning (Extreme Q-Learning), plus internships at Apple, Google AI and Uber ATG.[35]
Steve FreyCo-founder and President[1]The company's biography says he previously founded a fintech company, did research at Stanford and Berkeley, and did founding research work at Eon (connectomics) and Kernel (neurotechnology hardware).[1]
Naman GargCo-founder and CTO[1]First engineering hire at MultiOn, according to the company.[1] His website lists him as MultiOn's chief technical officer and founding member from August 2023, after work as a software engineer at Amazon Audible.[36] Co-author of the Agent Q and REAL papers.[13][5]

Expanded article table

The company lists Rich Miner, Richard Socher, Christopher Manning and Stefano Ermon as advisors.[1] Ermon is a co-author on several of Div Garg's papers, including IQ-Learn and Extreme Q-Learning.[35]

Products and technology

AGI, Inc.'s agents are computer-use agents: they act on what is shown on the screen instead of calling app-specific APIs or integrations, an approach the company calls computer-use or, on phones, mobile-use.[25][27][9]

AGI-0 models

AGI-0 is the company's name for its family of small action models. The company says it trains the models itself and that they range from 2 billion to more than 20 billion parameters. They read the screen, break a request into steps and then tap and type to complete it.[2][1] For its October 2025 AndroidWorld result, the company said its agent worked "entirely from screenshots," without accessibility trees, using a multimodal model fine-tuned for user-interface understanding. It added that the agent plans before acting and checks after each step whether the intended change happened.[20] Its MWC 2026 exhibitor listing called the models "Specialized Action Models (SAMs)" that use reinforcement learning to adapt to user and enterprise needs.[10]

AGI app

The consumer product is an assistant app called AGI, published on Google Play by AGI, Inc. Users speak or type a task, and the app carries it out through Android's Accessibility Service. As of October 2026 the listing showed more than 10,000 downloads and described the app as an assistive tool for people who find touchscreens hard to use.[37] The company's website lists Android as available, iOS as an early preview and PC as coming soon.[2] In a July 2026 interview, Garg described the system as a hybrid in which actions and light reasoning run on the phone while heavier reasoning runs in the cloud. He said purchases, messages and account changes require user approval.[38]

AGI-RUN

AGI-RUN is the company's inference engine for phones, laptops and edge hardware. In a July 2026 post, the company reported tests on a Samsung Galaxy S25 Ultra with six small vision-language models from the Qwen, Gemma and Liquid AI LFM families. It said AGI-RUN averaged about five times faster prefill than each model's native engine, up to 6.9 times. Against llama.cpp, MNN, MLC-LLM, LiteRT and mllm, it reported winning 58 of 60 model-and-workload cases, with a geometric mean of 3.3 times and a maximum of 9.9 times over the fastest rival. Google's LiteRT won the other two cases. The company said a full per-case table would follow in a later technical post.[31]

AGI MCP and platform

In August 2026 the company released an MCP server in developer beta that lets an MCP client such as Claude Code, Codex or Cursor operate apps on a connected Android phone. It said iOS support would follow, described uses including app testing on physical phones and control from smart glasses, and announced a planned voice SDK for app developers.[4] The company's Qualcomm announcement said its platform could also support agents such as the OpenClaw project.[25]

REAL benchmark and AGI SDK

REAL's replica sites include Omnizon (modeled on Amazon), Staynb (Airbnb), DashDish (DoorDash), GoMail (Gmail), NetworkIn (LinkedIn), UDriver (Uber), FlyUnified (United), TopWork (Upwork) and Zilloft (Zillow). Action tasks are graded by checking the site's state, and information-retrieval answers are graded by a rubric-guided language-model judge.[5] The paper reports that frontier language models achieved at most 41% success.[5] The company's launch post put Claude 3.7 Sonnet Thinking first among the tested models at 41.1%. Among agent frameworks, it reported 41% for Anthropic's computer-use agent, 31% for Browser Use, 19% for Stagehand and 7.1% for OpenAI's CUA, and said its own unreleased "AGI Agent-0" scored 45%.[8] The harness is distributed as the Apache-2.0-licensed agisdk package.[16]

Benchmark claims

The company's benchmark results are self-reported. The AndroidWorld leaderboard warns that its entries are community-submitted and not independently verified.[34]

BenchmarkDateClaimed resultNotes
REAL (company's own benchmark)April 202545% for internal "AGI Agent-0"[8]112 tasks on 11 replica websites[5]
AndroidWorld (116 tasks across 20 Android apps)[39]October 8, 202597.4%, screenshot-only[20]Company described it as a leaderboard record at the time[20]
AndroidWorldSeptember 2026100% (116 of 116 tasks)[2]Listed on the leaderboard as "AGI-0" with model "AGI-AW," pass@1, with the note "97.4 --> 100. Trained open source models using our novel training technique on AndroidWorld tasks." Four other entries also show 100% as of October 2026.[34]
OSWorld (369 computer tasks)[40]October 23, 202576.26% (274.52 of 360), versus a cited human baseline of about 72%; the company said training was "scaled with hundreds of thousands of synthetic and OSWorld tasks"[21]An archived November 2025 copy of the OSWorld-Verified leaderboard data lists an "agi-0" entry from AGI Inc at 65.42% with a 50-step limit.[41] The data published in October 2026 contains no AGI Inc entry.[42]

Expanded article table

Funding and investors

DateEventAmountValuationSource
June 2025First round, pre-seed/seed$8 millionNot disclosedEntrepreneur[3]
By November 2025Total raised, investors including Menlo VenturesClose to $10 millionNot disclosedForbes[7]
November 2025Reported talks (not confirmed as closed)$50 million sought$500 million (reported)Forbes[7]
April 2026Raising another roundNot disclosedNot disclosedEntrepreneur[3]

Expanded article table

TechCrunch described the company in January 2026 as "reportedly valued at $500 million" and backed by $10 million from Menlo Ventures, Anthropic and other investors. Its story carried a correction that the amount raised was $10 million, not $30 million.[9] AGI, Inc. lists Menlo Ventures and Point72 Ventures as investors. It also names angel investors including Charlie Songhurst, Gautham Thambidorai, Ali Rowghani, Songyee Yoon, Eugenia Kuyda, Nikita Shamgunov and Vijay Krishnan.[1] Its MWC 2026 exhibitor listing said: "Funded by Anthropic/Menlo Ventures, Point72, as well as angel investment from VPs at OpenAI, DeepMind, and Meta."[10]

Partnerships

PartnerDate announcedNature, as describedSource
VisaAugust 2025Partnership on agent-led payments and bookings ("agentic commerce")Company announcement[19]
QualcommMarch 2026Optimizing AGI's models and agent stack for Snapdragon devices, aiming at full on-device processingCompany announcement; Enterprise AI World[25][26]
LenovoMarch 2026Limited proof-of-concept inside Lenovo Qira at MWC 2026; not a product announcementCompany announcement with a Lenovo quote[27]
SamsungReported 2025-2026Forbes reported partnership discussions; TechCrunch said the company was working with partners including Qualcomm and SamsungForbes; TechCrunch[7][9]

Expanded article table

Reception

Forbes and the company itself have used the label "neolab" for AGI, Inc.[7][10] Debarghya Das's "The Ultimate Neolab List" includes the company but rates it "Borderline": it trains its own action models and has a researcher founder, but its raise is small and it is "product/consumer oriented."[43]

Tom's Guide called the beta app "impressively effective" in its April 2026 hands-on and compared the idea with the Rabbit R1. It also called the use of cloud GPUs "a little concerning from a privacy perspective."[29] Customer and usage figures come mostly from the company: TechCrunch reported 30,000 waitlist sign-ups in January 2026, and Garg told Entrepreneur in April 2026 that the number had reached about 500,000.[9][3]

See also

References

  1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19AGI, Inc., "About AGI, Inc.", company website, accessed October 27, 2026.
  2. ^1 ^2 ^3 ^4 ^5 ^6AGI, Inc., "AGI-0: small action model for on-device agents", company website, accessed October 27, 2026.
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Sherin Shibu, "OpenAI Offered Me a Million Dollar Job. I Said No. Here's What I Did Instead.", Entrepreneur, April 30, 2026.
  4. ^1 ^2 ^3 ^4Div Garg, "An MCP for every app on your phone", AGI, Inc., August 6, 2026.
  5. ^1 ^2 ^3 ^4 ^5 ^6Divyansh Garg, Shaun VanWeelden, Diego Caples, Andis Draguns, Nikil Ravi, Pranav Putta, Naman Garg, et al., "REAL: Benchmarking Autonomous Agents on Deterministic Simulations of Real Websites", arXiv:2504.11543, April 15, 2025 (v2: April 17, 2025).
  6. ^Div Garg et al., "REAL: Benchmarking Autonomous Agents on Deterministic Simulations of Real Websites", Advances in Neural Information Processing Systems 38 (NeurIPS 2025), Datasets and Benchmarks Track.
  7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Anna Tong, "AI Startup AGI, Inc In Talks To Raise Money At $500 Million Valuation", Forbes, November 19, 2025 (archived copy).
  8. ^1 ^2 ^3 ^4 ^5AGI, Inc., "Introducing REAL Bench: A New Standard for Web AI Agent Evaluation", April 18, 2025.
  9. ^1 ^2 ^3 ^4 ^5 ^6Sarah Perez, "AGI Inc.'s Android app can use your phone for you", TechCrunch, January 6, 2026.
  10. ^1 ^2 ^3 ^4 ^5MWC Barcelona, "AGI, Inc.", exhibitor listing (Hall 7, Stand 7G21), accessed October 27, 2026.
  11. ^AGI, Inc., "AGI, Inc.: Private action AI for phones and devices", company website, accessed October 27, 2026.
  12. ^U.S. Securities and Exchange Commission, "AGI Inc prospectus: 20,000,000 Class A Common Shares", Form 424(b)(1), filed February 11, 2026.
  13. ^1 ^2Pranav Putta, Edmund Mills, Naman Garg, Sumeet Motwani, Chelsea Finn, Divyansh Garg and Rafael Rafailov, "Agent Q: Advanced Reasoning and Learning for Autonomous AI Agents", arXiv:2408.07199, August 13, 2024.
  14. ^Div Garg (@divgarg), "Chapter II. Update: I will be moving on from @MultiOn_AI to explore new things!", X, December 20, 2024.
  15. ^MultiOn, multion.ai (redirects to theagi.company), accessed October 27, 2026.
  16. ^1 ^2AGI, Inc., "agi-inc/agisdk", GitHub repository (Apache-2.0), accessed October 27, 2026.
  17. ^DeepLearning.AI, "Building AI Browser Agents", short course taught by Div Garg and Naman Garg, 2025.
  18. ^Div Garg, "Everyday AGI: Democratizing Intelligence for an Abundant Future", AGI, Inc., May 1, 2025.
  19. ^1 ^2AGI, Inc., "AGI, Inc announces partnership with Visa to transform agentic commerce", August 16, 2025.
  20. ^1 ^2 ^3 ^4AGI Research Team, "The World's Most Capable Mobile-Use Agent", AGI, Inc., October 8, 2025.
  21. ^1 ^2AGI Research Team, "The World's Most Capable Computer Agent", AGI, Inc., October 23, 2025.
  22. ^Cade Metz, "Silicon Valley Builds Amazon and Gmail Copycats to Train A.I. Agents", The New York Times, December 2, 2025 (paywalled; archived copy).
  23. ^Techmeme, "Cade Metz / New York Times: A look at startups like AGI and Plato, which build replicas of websites to let AI agents learn to navigate the internet", December 3, 2025.
  24. ^Div Garg, "Apps Are Dead", AGI, Inc., January 1, 2026.
  25. ^1 ^2 ^3 ^4 ^5AGI, Inc., "AGI, Inc. announces Collaboration with Qualcomm Technologies to Bring Agentic AI to Snapdragon-Powered Devices", March 3, 2026.
  26. ^1 ^2"AGI, Inc. Partners with Qualcomm to Enable New Agentic Capabilities for Devices Powered by Snapdragon", Enterprise AI World, March 5, 2026.
  27. ^1 ^2 ^3AGI, Inc., "AGI Inc. Demonstrates Prototype Mobile-Use Agent Platform at MWC with Proof-of-Concept Lenovo Integration", March 11, 2026.
  28. ^Counterpoint Research, "Counterpoint Conversations: AGI on How On-device AI Agents Will Replace Apps on Your Smartphone", April 7, 2026.
  29. ^1 ^2Jason England, "I found the future of Siri in iOS 27, and it's currently an AGI Android app", Tom's Guide, April 5, 2026.
  30. ^Adam Awale, "Up to 96% Faster Processing-Near-Memory, Without Touching the Hardware", AGI, Inc., May 20, 2026.
  31. ^1 ^2Adam Awale, "AGI-RUN: Benchmarking AGI's on-device inference engine against the field, up to 9.9x faster than SOTA", AGI, Inc., July 2, 2026.
  32. ^Div Garg, "Hands-Free Phone Use: Building AGI With the Accessibility Community", AGI, Inc., July 31, 2026.
  33. ^AGI, Inc., "Rich Miner, co-founder of Android, joins AGI, Inc. as an advisor", August 4, 2026.
  34. ^1 ^2 ^3AndroidWorld, "AndroidWorld leaderboard", community-submitted results, accessed October 27, 2026.
  35. ^1 ^2Div Garg, "Div Garg", personal website, accessed October 27, 2026.
  36. ^Naman Garg, "Naman Garg", personal website, accessed October 27, 2026.
  37. ^Google Play, "AGI", app listing by AGI, Inc., accessed October 27, 2026.
  38. ^Antoine Tardif, "Div Garg, CEO and Co-Founder of AGI, Inc: Interview Series", Unite.AI, July 27, 2026.
  39. ^Christopher Rawles et al., "AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents", arXiv:2405.14573, May 23, 2024.
  40. ^Tianbao Xie et al., "OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments", arXiv:2404.07972, April 11, 2024.
  41. ^OSWorld, "osworld_verified_results.xlsx", OSWorld-Verified leaderboard data, Internet Archive copy of November 18, 2025.
  42. ^OSWorld, "OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments", leaderboard page and OSWorld-Verified results data, accessed October 27, 2026.
  43. ^Debarghya Das, "The Ultimate Neolab List", accessed October 27, 2026.

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

1 revision · v1 · 3,771 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Founders, funding reports, products, partnerships and self-reported benchmark results checked against company posts, Forbes, TechCrunch, Entrepreneur, The New York Times, the AndroidWorld and OSWorld leaderboards and the REAL paper, October 2026.

Cite this page: AI Wiki. "AGI, Inc.." aiwiki.ai, updated 27 Oct 2026, fact-checked 27 Oct 2026. CC BY 4.0. https://aiwiki.ai/wiki/agi_inc

Suggest edit

Where to go next

Topics:AI AgentsAI CompaniesAI Research