IrisGo Wants to Be an AI Desktop Agent That Learns by Watching
IrisGo, a startup backed by Andrew Ng's AI Fund, is building a desktop AI agent that watches what you do on your computer and learns to automate those tasks for you. The company, which originally pitched itself as an 'AI butler,' is positioning Iris as a persistent, observational productivity layer.
Original sourceIrisGo is entering the increasingly crowded desktop AI agent space with a product that takes a passive-first approach: rather than asking users to define automations upfront, Iris sits in the background, watches the user's desktop activity, and infers which repetitive tasks it should learn to handle. The startup is backed by Andrew Ng's AI Fund, lending it early credibility in a category that's seen a lot of vaporware.
The core pitch is behavioral learning over explicit configuration. Instead of scripting macros or building workflows in a no-code editor, a user theoretically just works, and Iris builds a model of their habits over time. Co-founder claims range from automating multi-step copy-paste workflows to handling routine email triage and file organization — tasks that currently require either manual effort or significant setup time in tools like Zapier or Keyboard Maestro.
The privacy implications of a tool that continuously observes your screen are significant and largely unaddressed in current public materials. IrisGo will need to answer hard questions about where screen data is processed, how long it's retained, and what model training looks like, especially for enterprise adoption. The 'butler' framing has been softened, but the fundamental product — an always-on screen watcher — hasn't changed.
Andrew Ng's backing puts IrisGo in notable company alongside other AI Fund bets, and brings distribution credibility in enterprise and developer communities. But the gap between 'watches your desktop and learns' as a demo and 'watches your desktop and reliably automates' as a shipped product is where most of this category goes to die. IrisGo's ability to close that gap — and to do it without becoming a privacy liability — will determine whether this is a product or a pitch.
Panel Takes
The Skeptic
Reality Check
“This is Recall by Microsoft, but smaller and with Andrew Ng's name on it — and Microsoft had to pause Recall because users revolted over the privacy implications of continuous screen capture. IrisGo hasn't published a single concrete answer about where your screen data goes, how it's processed, or whether it's used for model training, which means the product is currently a demo with a privacy time bomb attached. The specific scenario where this breaks: any user at a company with a security policy, which is most of the users who'd actually pay for this.”
The Builder
Developer Perspective
“The primitive here is 'screen-capture-to-task-model pipeline with an inference layer on top,' and the DX bet is zero-configuration learning — complexity is pushed entirely to the runtime instead of the setup. That's the right call philosophically, but I want to see the API surface: can I inspect what Iris thinks it learned, correct it, or compose its automations with other tools, or is it a black box that just acts? If there's no way to introspect or override the learned model, this isn't a productivity primitive — it's a liability you can't debug.”
The Founder
Business & Market
“The buyer for this is either a power user paying out of pocket or an IT department, and those two buyers have completely opposite requirements — one wants it to just work, the other needs audit logs, access controls, and a data processing agreement. Andrew Ng's backing is real signal, but the moat question is brutal: the moment Cursor, Notion, or any OS-level player ships 'learn from watching,' IrisGo's core differentiator evaporates unless they've built proprietary behavioral models that compound with usage data. The business survives only if the learned automation graph becomes sticky enough that switching costs are real — right now there's no evidence that's been designed in.”
The Futurist
Big Picture
“IrisGo's thesis is that the right interface for automation is observation, not instruction — that users will never reliably describe their own workflows but will always perform them, and that the delta between 'watched' and 'automated' will close fast enough to make this viable before the OS vendors get there. That's a real bet, and it's early on the trend line of ambient compute agents, but the dependency chain is long: it requires on-device model inference good enough to generalize from observed behavior, privacy-preserving architecture that survives enterprise scrutiny, and a learning loop tight enough that automation accuracy compounds rather than plateaus. If Apple or Microsoft ships this at the OS level — which they've both signaled — IrisGo needs to have built something those platforms can't replicate in 18 months, and nothing in the current public materials suggests what that is.”