There were 2 Tsunamis that hit the AI world last week - DeepSeek and OpenAI's Operator. This post is about Operator and before I get into what Operator, it helps to set some context. I've spent the last year going all in on GenAI.
I told myself 1 thing - "GenAI is going to be a major disruption and I had a unique opportunity (I had gotten let go from my job and I had the freedom, family, time and financial, to do whatever I wanted)".
I started with gaining expertise on Assistants and Co-Pilots then Prompt Engineering and then building Agents using low code / no code tools like AnyQuest.ai, Clay.com and Agent.ai (to name a few). I tried out 50 other such tools.
It is amazing how fast the world changes. A year ago it was all about Assistants/CoPilots and Prompt Engineering. Microsoft was high on Office365 CoPilot, the champagne was popped. Then the rumblings started around people not getting value from CoPilots and by the end of the year it was all about Agents.
“I have yet to find anyone who's had a transformational experience with Microsoft Copilot or the pursuit of training and retraining custom LLMs” the Salesforce CEO concluded. “Copilot is more like Clippy 2.0.
So when Operator came out last week, I was very curious if this was going to be the future of agent building. Would my expertise (if you can call 6 months as that) in prompt engineering and building agents be redundant. Would Operator just build autonomous agents without the need for an agent builder.
So I've spent this weekend getting hands on with Operator. I decided to take some simple agents I've built with Agent.ai, AnyQuest.ai and Clay.com and see if I can replicate them with Operator. I created a YouTube Playlist to record my learnings and I will be posting more examples of my experience.
I started by bulding 3 simple agents -
Takes the name of a person and uses their LinkedIn data to generate their DISC profile (this is one of my more popular agents on Agent.ai) - DISC Profile Agent.
My car broke down and I wanted to build an agent that would find local automotive shops within 25 miles, pull their contact info and reviews and give me the top X. I've built something like this with Clay.com and their Google Maps integration.
I suck at writing good titles / descriptions / tags for my YouTube videos. So I built a simple agent that takes a YouTube video and will generate these automatically (I also built an agent that can loop through all my YouTube videos and do this at scale and update YouTube). So I needed this agent to generate the output as JSON. - YouTube Metadata Agent
For those interested in watching the playlist as I build these using Operator, here is the link - Will Operator Take My Job
So what did I find out -
So this was my first iteration with Operator. The concept of it mimicking human clicks felt kluge. It definitely does not seem to be the use case where I would use this against large lists of items (e.g. research these 50 people) unless it is a long running process and I could just let it run for a day. Interacting with a website or data via API is much faster and more intuitive than interacting with a website via clicks. For simple tasks like reading someone's LinkedIn profile and summarizing it, it works fine.
Just as I was getting ready to write it off as something cute and not really something I could see using in production, I thought, why not give it a crazy task. What if I asked it to log in to SaaS software I use (Agent.ai, Anyquest.ai, Clay.com, Tableau Public etc.) and build an agent or build a dashboard or build a Clay table. This seemed like a ridiculous ask but I was curious and this is where things got very interesting.
It actually was able to take some simple instructions I gave and log in to the software and use it. This was crazy as it looks like it is learning the software on the fly by clicking through it. It didn't seem like it was reading the product manual or reading blogs. I am very very curious to get deeper into how it is doing this. So here were the instructions I gave it -
To build the Tableau Dashboard -
Can you log into Tableau Public and build a quick dashboard that show orders by product and region using the Super store data set
The dataset Superstore can be found in one of these chat threads here.. can you get it from there - https://community.tableau.com/s/question/0D58b0000CFiCI1CQN/where-is-most-current-super-store-dataset-using-tableau-public
Can you go to the Connectors tab and connect to my Google Drive
This was the output - https://operator.chatgpt.com/v/6796be1d3b6c819287af1fff6cd785e0
To build the Agent.ai DISC Agent -
I'd like you to log in to Agent.ai and build an agent that takes a user's input of a name of an executive and the company they work at and then it finds their LinkedIn URLs and then uses that to load their LinkedIn profile and posts and then uses that to generate a DISC profile
This was the output -
https://operator.chatgpt.com/v/6796afad94a481929acbaf1e803ea052
Here is a video where I capture my thoughts on this and all the videos -
So where did this experience leave me - I do think at some point Operator will build Agents on Agent.ai or build a dashboard on Tableau. Maybe simple ones, who knows. That being said, there will definitely be a need for someone to 'train' Operator to do this well and I don't know what 'train' means. Is it just instructions? Is it giving it examples and training it? Will I feel like I am training my replacement? I don't know.
What did stand out for me was that I have for the longest time struggled with finding a good use case of GenAI for Professional Services and Customer Success and I can see the potential of using Operator to do some elements of Onboarding / Set up / Training etc.. I think these functions will be disrupted in some ways we are not even aware of.
So yes, it might take my current job but it will create new jobs!!