MultiOn
Last reviewed
Sources
18 citations
Review status
Source-backed
Revision
v1 · 2,180 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
18 citations
Review status
Source-backed
Revision
v1 · 2,180 words
Add missing citations, update stale details, or suggest a clearer explanation.
MultiOn (stylized MULTI·ON, also written MultiOn AI) is an American startup that builds autonomous web-browsing AI agents, software that can read web pages and take actions on a user's behalf such as booking restaurants, ordering food, filling forms, and shopping. Founded in 2022 by Div Garg and Omar Shaya, who met at Stanford University, the company released an early browser extension in January 2023, went viral in late 2023 for demos of an agent autonomously booking flights and making reservations, and emerged from stealth in January 2024. MultiOn is best known for two things: an Agent API that lets developers embed web-automation agents into their own software, and Agent Q, a 2024 research paper on self-improving web agents co-authored with Stanford researchers. The company has since reorganized under the name AGI, Inc. and pivoted toward a mobile-first consumer assistant; co-founder and chief executive Div Garg stepped back from day-to-day leadership in late 2024.
MultiOn grew out of Stanford in 2022. According to Amazon's Alexa Fund, Omar Shaya conceived the idea in March 2022 while studying business management at Stanford, then met Div Garg, who was pursuing a Ph.D. in artificial intelligence and reinforcement learning at the university, and the two began collaborating in July 2022. Several profiles describe the pair as having met through Garg's Stanford seminar CS 25: Transformers United, a course on the transformer architecture that Garg created and that has drawn more than a million views on YouTube.
Div Garg (full name Divyansh Garg) took leave from his Stanford AI Ph.D. to build the company and served as chief executive. Before Stanford he earned a master's in computer science from Cornell, where he worked on computer vision and early camera-based autonomous-driving systems, and his research spanned reinforcement learning, imitation learning, and language-controlled agents. Omar Shaya, the business co-founder, had a product-management background that included stints associated with Meta and Microsoft.
A third central figure is Naman Garg, who joined in August 2023 and is described on his own profile as MultiOn's "Chief Technical Officer & Founding Member." He led much of the engineering: an advanced remote browser agent, the company's "Large Action Model" (LAM), the MultiOn iOS app, the browser extension, and backend APIs. Naman Garg is also a co-author of the Agent Q paper and later co-founded the successor venture AGI, Inc. with Div Garg. (Some secondary write-ups loosely refer to "Div and Naman Garg" as the founders, but the founding pair documented by Amazon and contemporaneous interviews is Div Garg and Omar Shaya, with Naman Garg joining as a founding engineer in 2023.)
MultiOn's first limited product, a browser-based extension, launched in January 2023. The company operated semi-quietly through 2023, building a waitlist that it said exceeded 30,000 people by November 2023, before formally coming out of stealth in January 2024 around the CES 2024 trade show, where its technology was shown in Amazon's exhibit space.
MultiOn's funding has been reported but never accompanied by a detailed, fully itemized public disclosure from the company, so amounts and round labels vary across sources. The most widely cited event, reported by The Information and echoed by multiple outlets, is a roughly $20 million raise at about a $100 million valuation announced around the company's January 2024 emergence from stealth, led by General Catalyst. Sources disagree on whether to call this a seed or a Series A; Crunchbase logs a round dated January 23, 2024, while some aggregators label a $20 million "Series A" in mid-2024. Coverage at the time framed the valuation as a steep climb roughly twelve months after founding and less than six months after the first product announcement.
Reported investors across the seed and the 2024 round include General Catalyst as lead, Amazon's Alexa Fund, Samsung Next, Maven Ventures, and angel investors such as Nat Friedman and Daniel Gross, along with what the company described as individuals from OpenAI and early backers of DeepMind. Amazon's Alexa Fund publicly confirmed MultiOn as a portfolio company in January 2024.
| Round | Amount (reported) | Valuation (reported) | Lead | Date reported | Selected other investors |
|---|---|---|---|---|---|
| Seed | undisclosed | not disclosed | General Catalyst | 2023 | Amazon Alexa Fund, Samsung Next, Maven Ventures, Nat Friedman, Daniel Gross |
| Round labeled seed or Series A | ~$20 million | ~$100 million | General Catalyst | January 2024 (per The Information) | Amazon Alexa Fund, Samsung Next, angels |
The table reflects how the funding was reported in press coverage; the company has not published an authoritative round-by-round breakdown, and some third-party databases list higher cumulative totals that are not corroborated by primary sources.
Through 2024 MultiOn shifted its emphasis from a broad consumer agent toward developer APIs, enterprise partnerships, and narrow "vertical" agents for specific tasks such as restaurant reservations, travel, and scheduling. In a December 3, 2024 interview, Div Garg framed the strategy around domain-specific agents built in partnership with businesses rather than a single open-ended consumer product.
On December 21, 2024, Garg announced on X that he would move on from MultiOn to "explore new things," remaining as an advisor to the team. He went on to found AGI, Inc. (also styled AGI Inc. and operating at theagi.company), which he describes as an applied research lab focused on agent reliability, trust, and human-AI collaboration, with a small team of "serial founders and seasoned engineers." Naman Garg is a co-founder of AGI, Inc.
The MultiOn brand and the AGI, Inc. effort have become intertwined. The company's old multion.ai blog now redirects to theagi.company, and the consumer product has been relaunched under the AGI, Inc. umbrella as a mobile assistant (reported under names including "Please" and the AGI-0 app), reflecting a mobile-first pivot. As of 2026 the operation is small (third-party trackers list a single-digit headcount), and the company has been comparatively quiet relative to its high-visibility 2023-2024 period.
The original MultiOn product was a personal AI agent delivered first as a browser extension (with support reported for Chrome, Edge, and Arc) and later as a mobile app. The agent accepts natural-language instructions and then navigates and clicks through websites to complete multi-step tasks. The late-2023 demos that drew attention showed the agent ordering food on DoorDash, booking flights, and making dinner reservations. The company reported "10,000+" beta users and a sharp month-over-month usage spike in a February 2024 interview, figures it stated itself and which were not independently audited.
MultiOn's developer-facing Agent API lets software issue natural-language commands that an agent carries out in a browser. The API was first announced around August 2023 and opened in public beta in 2024. According to MultiOn's documentation, the core capabilities include:
| Capability | What it does |
|---|---|
| Browse | Executes a natural-language goal, taking multi-step actions on a site until the task is done or input is needed |
| Retrieve | LLM-powered structured data extraction (scraping) that returns specified fields as JSON, with options for full-page crawling, JavaScript rendering, and scrolling |
| Sessions | Stateful interactions: create a session, step through it with instructions, fetch screenshots, list or close sessions (sessions persist through about 10 minutes of inactivity) |
The platform offers two execution modes: cloud-hosted remote sessions that run a headless browser with native proxy support to get past bot protection and IP blocks, and a local mode in which the agent drives a user's own installed Chrome via the MultiOn browser extension. Official SDKs are provided for Python and for JavaScript/TypeScript (Node 18+, with compatibility noted for Vercel, Cloudflare Workers, Deno, and Bun), both featuring automatic retry with exponential backoff on retriable HTTP statuses. The documentation references an "Agent V1 Beta" with an "Agent 2.0" planned to add Agent Q support. MultiOn positioned the API for use with developer toolchains and it has been integrated into popular orchestration libraries.
Rather than train its own foundation model, MultiOn customizes open-source large and small language models to control applications, an approach it markets as a "Large Action Model" (LAM). The 2024 Agent Q research used Meta's Llama 3 70B as a base model, illustrating the company's strategy of fine-tuning existing large language models for agentic web tasks.
Agent Q is MultiOn's most cited technical contribution: a framework for autonomous, self-improving web agents published as the paper "Agent Q: Advanced Reasoning and Learning for Autonomous AI Agents" (arXiv:2408.07199), submitted August 13, 2024, with an accompanying announcement around August 13 to 14, 2024. The work was done by MultiOn (credited in the paper as "The AGI Company (MultiOn)") together with Stanford University. The authors are Pranav Putta, Edmund Mills, Naman Garg, Sumeet Motwani, Chelsea Finn, Divyansh Garg, and Rafael Rafailov; the listed corresponding contacts are Div Garg (MultiOn) and Rafael Rafailov (Stanford). Chelsea Finn and Rafael Rafailov, both at Stanford, are the academic co-authors. (The name "Agent Q" plays on speculation about OpenAI's then-rumored "Q*" project, though the two are unrelated.)
Agent Q combines three ideas to let a web agent improve from its own experience with limited human supervision:
The paper reports two main evaluations. On the WebShop benchmark, an agentic shopping environment, the base xLAM model scored about 28.6% success, which Agent Q with MCTS raised to roughly 50.5%, surpassing the paper's reported average human performance of about 50%. In real-world restaurant-booking experiments on OpenTable, MultiOn reported that Llama 3 70B improved from an 18.6% zero-shot success rate to 81.7% after a single day of autonomous data collection (which the company described as a roughly 340% relative jump), and further to 95.4% when online search was added at inference time. These figures come from MultiOn's own paper and announcement; the OpenTable result in particular is a self-reported, real-world experiment rather than an externally administered benchmark.
MultiOn was among the more visible early entrants in the autonomous web-agent space in 2023 and 2024, alongside companies and projects working on similar "computer use" and browser-automation agents. The viral 2023 demos and the 2024 funding at a nine-figure valuation drew significant industry attention, and Agent Q was widely covered in the AI press as a notable result in agent self-improvement. The space it helped popularize subsequently saw entrants such as OpenAI's Operator and a number of browser-automation startups, several of which later reported higher scores on public web-agent benchmarks. MultiOn itself grew quieter after its 2024 leadership change and reorganization, and independent reporting on its current scale and metrics is limited, so claims about user counts and performance should be read as the company's own statements unless otherwise sourced.