Building an Agent with Operator
Our verdict is...
In our previous post, we explored the concept of building an agent that could fetch and summarize news, replicating the core functionality of our AI News Scout. This time, we decided to test OpenAI’s Operator, a tool introduced in January 2025 and updated in May.
Operator stands out because it’s built with reinforcement learning, enabling it to perform complex, multi-step tasks. In practice, that means it can click, scroll, and type as it navigates websites, like an AI assistant that can do your online grocery shopping or set up alerts for you.
Reinforcement learning:
An AI technique where the computer learns by trying out and getting rewards or penalties, like playing a game and earning points.
ChatGPT also uses reinforcement learning (as you’ll notice from the thumbs-up or thumbs-down under each response), but with Operator you can provide explicit feedback, a sentence or two explaining success or failure. That makes it a more goal-directed tool, trained not just on response quality but on task completion.
What Makes Operator Unique
Operator tackles multi-step workflows, adapting its actions based on feedback and outcomes. Unlike traditional AI models that respond to single prompts, Operator can:
Log into accounts
Set up Google Alerts
Fill out forms
Collect and organize data
This is a leap forward for consumer AI, aiming to automate routines that used to require manual effort.
The First Surprise: Price
The first thing to note: it’s expensive. Operator is bundled into the $200/month Pro package, which includes several other features, but you can't currently subscribe to Operator alone at a lower price. The Plus Plan ($20/month) doesn’t include it either. See below:
For most users, this high price will be a barrier. A recent study found only 3% of U.S. adults pay for premium AI services, especially if their needs are limited to what Operator offers.
Getting Started: Familiar Yet Different
At first glance, Operator looks a lot like ChatGPT. The interface is nearly identical, and prompting feels very familiar.
We asked ChatGPT to generate a prompt for our AI News Scout, and it delivered an impressively detailed workflow. Initially, Operator executed well, creating alerts and following a clear, logical path.
But hiccups surfaced quickly:
Workflows stalled or failed to complete
Some steps needed manual fixes or debugging
That led us to rethink the approach and start over…
Second Attempt: Simpler Is Better
For round two, we simplified the project, breaking it into smaller steps. See below the revised prompt:
“Create a daily report compiling the top 10 most relevant articles about consumer AI technologies and applications, focusing on new businesses, use cases, and forward-looking insights. Use reputable sources like TechCrunch, MIT Technology Review, The Verge, Wired, Bloomberg, Financial Times, Forbes, CB Insights, Ben’s Bites, Import AI, Sequoia, a16z, and more. Rank articles by application-layer focus, consumer relevance, novelty, and accessibility. Format as a table with title, source, date, score, and summary. Send daily to xxxx@proton.me via ProtonMail.”
This time, it worked better. Operator took a while to find the news, email login was clunky but it allowed for some hands-on control.
Unexpected Frictions
A few frustrating moments worth calling out:
Google blocks logins from virtual browsers, like Operator’s. You’ll need a workaround, like using ProtonMail.
To send emails automatically, you’ll need to set up integrations with tools like Zapier, which may be challenging for non-technical users (me).
After creating the email to be sent, the format was dreadful and re-formatting was a battle that I lost miserably.
Once the task is saved, rerunning it doesn’t just fetch new output. It re-executes the entire flow (e.g., re-setting alerts), which gets repetitive and time consuming!
The Verdict: Not Worth It—Yet
While Operator has potential, its current value is hard to justify, especially with a $200/month price tag. For most users, manually setting up alerts and using tools like Perplexity for daily summaries is far more practical.
Here’s our own test using the same prompt in Perplexity (here ). The result? A concise, well-formatted summary stored in “My Spaces.” It took seconds, cost nothing, and while it didn’t email me directly, I’d rather receive that than a messy, unformatted alert.
Looking Ahead
Will OpenAI keep investing in Operator? Will better use cases or lower pricing make it more compelling? Time will tell.
We’ll keep experimenting for sure! and if you have experience with Operator, let us know in the comments. Your insights could help shape our next test.
For now, though... we’re leaning toward cancelling our subscription 😊
Stay tuned as we continue exploring the evolving consumer AI ecosystem.
Previous blogs