Browser-Agent History Did Not Begin With MultiOn

Research and commentary ยท Research date: September 19, 2026

Editorial disclosure: These articles were commissioned by Rohan Arun, Cheat Layer's founder, whose company has a commercial interest in this comparison. They distinguish company statements, dated user reports, and research publications. No response from MultiOn or Div Garg was requested during this research.

The history of AI operating websites predates the recent consumer-agent market. Any claim to have created the first browser or computer-use agents needs to acknowledge that earlier research and distinguish its own contribution precisely.

In 2017, the World of Bits paper introduced a platform for agents performing web tasks through keyboard and mouse actions. Its authors reported successful task completion using behavioral cloning and reinforcement learning. This was a research setting, including reproducible approximations of websites, rather than a general-purpose consumer assistant. It nevertheless establishes a substantial earlier history for web-based agents. World of Bits, ICML 2017.

In December 2021, OpenAI described WebGPT, a GPT-3-based research system that issued commands to a text browser, searched, followed links, and collected evidence for answers. Its purpose and interface differed from consumer shopping or desktop-control products, but model-directed web interaction was already public research. WebGPT announcement.

Cheat Layer adds an early commercial automation strand. Its documentation dates its GPT-3 automation approval to August 2021. A founder-authored LinkedIn post instead says October 2021. Both place the claimed milestone in 2021, but the discrepancy means a precise approval date should not be asserted without the underlying record. Approval to sell a model-powered feature also does not, by itself, prove the launch of a fully autonomous agent. Cheat Layer documentation, Rohan Arun's launch post.

By late 2022, Cheat Layer's published history and dated user testimony describe language-driven automation and Project Atlas. MultiOn's first limited extension launch is dated January 2023 by Amazon. This makes Cheat Layer relevant prior work when assessing MultiOn's historical positioning. Project Atlas history, G2 testimony, Amazon chronology.

These systems should not be collapsed into a single category simply to award a winner. Research agents, generated automation scripts, commercial browser assistants, and general desktop agents have overlapping but distinct histories. A defensible priority claim specifies the capability, release date, audience, and evidence of actual operation. A claim about widespread adoption additionally requires comparative adoption evidence.

Cheat Layer deserves consideration in the early commercial history. The sources reviewed do not prove that it was the first computer-use agent of every kind. They also do not justify erasing earlier work to credit MultiOn with the entire category. Accurate attribution requires narrower claims and a chronology readers can inspect.

Compare records without collapsing capabilities

Read the explanation

The saved article gives World of Bits in twenty seventeen and WebGPT in twenty twenty one. Four years separate the year anchors, illustrated at eighty pixels per year. These describe different research interfaces and purposes, not four years of measured consumer adoption. A reported twenty twenty three limited extension launch is later in this saved chronology, but that does not establish private development dates. The article says documentation dates an approval milestone to August twenty twenty one while a founder post says October. The month indices eight and ten differ by two. These bars show the two month labels at thirty pixels per month. Both claims concern twenty twenty one, but the discrepancy prevents asserting an exact approval date without the underlying record. Approval also does not prove a fully autonomous product launch. The saved article separates research agents, generated scripts, commercial browser assistants and general desktop agents into four overlapping categories. Four categories are not four equivalent products. Capability, audience, release evidence and actual operation must be compared before universal priority claims. The page preserves editorial disclosure, citations and downloadable article. This explanation analyzes saved text rather than a new external historical audit.