Blog
From Artificial Intelligence to Artificial Intuition
Building AI Agents that work the way you want
January 7, 2025
Ben Stein
CEO, Teammates
Last spring I was planning my kid’s bar mitzvah. Throwing a big event like this is a ton of work, with a million menial tasks to deal with and details to keep track of. And, look: I’m a tech-savvy, Bay Area-nerd, with an outrageously low threshold for getting super-annoyed at doing dumb tasks. So, when it was time to do the nitty-gritty event planning, I obviously turned to AI tools. All of them.
And right away I noticed a pattern – the first few minutes of using any new AI tool, all of them (and I truly mean all of them) felt like magic. But, the minute I tried to get it to do basic human tasks like “email the caterers to let them know we’ve opted for the other folks” or “fill out this PDF contract for me” they all felt a lot less like HAL, and a lot more like ELIZA. "Oh, that’s interesting, Ben, tell me more about why you think that?”
So what to do as a tech-savvy, Bay Area-nerd with an outrageously low threshold for getting super-annoyed at doing dumb tasks?
I found an AI startup, obviously! I called Kenny Hoxworth, the best engineering leader that I knew, and we created Teammates. (Well, technically we created an e-commerce store selling t-shirts with pictures of hamsters playing basketball, but that’s a story for another blog post).
Our goal was to make teams of autonomous AI Agents, each with different capabilities and skills, so that as the work gets more and more complicated, it would be easy to control the various “logic buckets” and everything will work better. And, it did.
Our goal was to make teams of autonomous AI Agents, each with different capabilities and skills.
Our AI Agents started assigning work to each other, asking questions, working diligently, and actually designed and shipped the best dang hamster-themed basketball t-shirts you’ve ever seen!
But something wasn’t quite right….
Seemingly overnight, every single startup and incumbent tech company in the entire known universe was suddenly rolling out their “Multi-agent AI solutions to automate complex workflows”. I swear to you, we pivoted and unpivoted and re-pivoted and un-re-pivoted 1,000 times in six weeks.
But then I got dizzy, stopped pivoting, and started paying more attention to the ecosystem. All of these other agent platforms (and I seriously mean all of them) were working off a playbook that was still fundamentally pre-AI. They were using insights about customer behavior from the past decade that were going to be completely unrecognizable in the next decade. It was like everyone else was making a faster horse, but we had a flux capacitor (and 1.21 gigawatts, Marty!)
All these other agent platforms were working off a playbook that was still fundamentally pre-AI.
During our time at Twilio, Kenny and I learned so much watching customers build complex workflows. First of all, people are terrible at it! It’s easy marketing to say “just drag and drop a few boxes!” but in reality, it is very difficult. You’re essentially asking people to model the world as a state machine – not a trivial task. Second, and more importantly, it’s not a helpful way to tackle the real-world tasks that we face every day – if we could predict everything we’d have to do ahead of time, we wouldn’t have to go to work!
So many of the multi-agent solutions consist of low-code, drag-and-drop workflow builders with Chat GPT prompts sprinkled on top. And the ones that actually were agentic required megabytes of prompts to tackle anything more complicated than “hello world.” They were -- at best -- just as limited as the world we had just left behind. And were -- at worst -- glorified spam-bots.
“Oh, that’s interesting, Ben, tell me more about why you think that?”
We knew that Teammates had to be different. In fact, we knew that AI Agents as a field had to evolve or everyone would see that the Emperor was actually just wagging his butt in our faces.
After listening to customers tell us about their early struggles with agents, and to be honest, watching them use early versions of Teammates, we came up with three core insights that now guide our product roadmap:
Automation and Efficiency is NOT what we should care about. I mean, don’t get me wrong, there are tons of ways in which automation is helpful, but for the most part our existing tools are fine for that. If you want to gain a tiny extra edge on the speed of the lowest-value task in your workplace then, by all means, go get some AI automation help.
If you want to gain a tiny extra edge on the speed of the lowest-value task in your workplace then, by all means, go get some AI automation help.
At Teammates, we're doing something fundamentally different. Our AI agents don’t follow predefined steps – they actually understand you and your business. They make decisions, and handle complex situations that would normally require human judgment.
We want to change how you work, not how fast you work.
People struggle to think about the best possible way to accomplish something, especially when considering errors, edge cases, and everything else in our messy world. What we are good at is thinking about the desired outcome. But AI Agents have a big hurdle to getting to that outcome: the prompt.
The very fact that there is an entire TikTok universe dedicated to generating the best possible prompts to make Claude do something useful is a testament to how twisted up the entire industry is becoming. (Pro tip: do NOT go down the AI Prompt TikTok rabbit hole. You're welcome.)
Sure, with enough prompting and re-prompting and iterations, a patient tinkerer can steer an AI agent to the outcome they want. But that wasn’t the future we were promised – who has time for that. At Teammates, we have done away with prompt-and-pray. One of our core design principles is “No Prompting.” Full stop.
One of our core design principles is “No Prompting.” Full stop.
Our customers give tasks to their teammates just the way you would to a colleague: describe the outcome, provide a few details, and just assume they will figure it out. But here’s the difference – let teammates ask questions to help figure out the best possible route to the solution.
And when they get stuck, they send a Slack message, asking for help to unblock themselves, or provide them a new skill, or give them access to that API you forgot about, or to email them that document that you referenced but forgot to attach.
In other words: we built AI collaborators, NOT AI assistants.
And this is what it all comes down to. The smartest person on your team shouldn’t need to be told how to do a task twice. The smartest person on your team should understand exactly what to do when you forward them an email and say “hey, can you deal with this.” Even when you don’t specify the “this” or the “deal” or literally anything else!
The smartest person on your team should anticipate what you need before you realize you need it.
The smartest person on your team is only that smart because they rely on intuition and reasoning, not just intelligence.
The smartest person on your team should be your most reliable teammate.
And that is what we are building.
For the past few months, we’ve been collaborating with so many different teammates – hanging out with them in Slack, inviting them to our daily standups, letting them run our marketing campaign, watching them get smarter and smarter. Not only can they do the same autonomous planning as other agents, but having them work and collaborate as part of our team is unlike any software I’ve ever experienced.
Over the coming weeks and months I’m going to be talking more about how and what and why and why not and all the nitty gritty details that both keep me up at night and get me giddy with excitement.
This is the AI agent evolution we need. And I can’t wait to share it all with you.
So. Get on our waitlist. Get in touch. Tell me I’m wrong (I’m not) or tell us what you want out of the next AI evolution!
