People are not taking AI agents seriously enough.
With the recent release of OpenAI’s consumer-targeted AI agent (called Operator) it’s become clear that the intent of agentic models is to wholesale replace human work done on computers.
First, they’re coming for online work.
Operator functions best when navigating an internet browser. OpenAI’s marketing shows casual use, like having the agent order groceries or book a reservation for you.
But agents can go far beyond simple food orders. These models have been trained to replicate virtually anything a human can currently do in an internet browser. Communicate with your coworkers on Slack, develop and deploy a prototype, monitor and manage your stocks, the list goes on… and on… and on.
Goodbye, humans
Over 92% of jobs require digital skills. Whether they are in an office or sitting at home, a significant percentage of humans make a living exclusively from computer work. Developers, day traders, product managers, you name it. Industries of all types are saturated with humans clicking around various websites and performing tasks.
AI agents were made to click around on various websites and perform tasks. And right now, today, they’re pretty damn good at it. Take a look (thanks Yana Welinder):
In this example, Operator does the following within 5 minutes:
- Sets up an integration between Slack and a product management app (Kraftful)
- Prioritizes feedback and added to roadmap
- Writes a PRD based on user feedback, starts making user stories
- Syncs a user story (with additional info) to an issue tracking platform (Linear)
- Assigns the Linear ticket to an engineer
This is a product manager’s full flow for ticket creation, automated entirely. The possibilities for automation work with agents is near endless. It can:
- Write detailed notes from a meeting transcript, email out personalized action items to each attendee
- Create new recipes, write human-sounding recipe blogposts, and continually publish them to a food blog
- Analyze a competitor’s website, develop a version of the site, and deploy the site to Cloudflare or other hosting platform
- Monitor incoming messages on a customer service platform and reach out to users as needed
- Research a competitor’s marketing material and plan a week’s worth of marketing content
- Search through thousands of documents and databases on stock market data, identify the optimal moves, and make transactions every minute
Initially, the models powering agentic tools will be imperfect. Operator is slow and reaches out to its human handler often for assistance. But this will change rapidly as both the frontend agent product and the model powering it inevitably get more and more powerful.