This as-told-to essay is based on a conversation with Andy Cabasso, a growth operations manager at productivity platform ClickUp. He is based in New York. The following has been edited for length and clarity.
I used to practice law, and I became very disenchanted with the inefficiency of it, particularly the focus on billable hours. It didn’t benefit me to be better with keyboard shortcuts or to be able to work faster than my colleagues.
So I started a legal digital agency focused on leveraging automation with projects that were value-priced, rather than priced on inputs. I was acquired by a larger company, and now I focus on growth operations at ClickUp.
When I arrived at this company over a year ago, I wasn’t working with any agents. Then ClickUp gave us these tools and a tech stack to give us agents that had context on all of our work, and I thought, “How can I best use this?”
It snowballed from there.
I have 37 agents now
I used to say that I was the most technical non-technical person on my team. I didn’t know how to code anything — and I still don’t think anyone would pay me to code today.
But I understand a lot of the frameworks of how software actually works from working alongside technical cofounders and building marketing automation systems. I understand how databases are structured, how APIs connect, and how data flows between tools.
When I first started at ClickUp, we had a mandate to use AI agents more in our day-to-day. There was a culture fostered around sharing our AI workflows. I started small with the tasks that were tedious that I wanted to take off my plate. The first agent I built was a scheduling agent that would look at my calendar and task list, and help me plan out my day.
I have 37 AI agents now. I probably use 15 to 20 of them in my day-to-day, and the rest on an as-needed basis.
I have one agent that pulls analytics from our outbound platform to give insights into how our different campaigns are performing. It also proactively analyzes the campaigns, compares them against each other, provides us with recommendations on changes, and implements those changes.
I have an agent that reviews meeting transcripts. I’ve got another agent that will create follow-up tasks.
I don’t know what things will look like in a few years, but I will likely create more agents and will have more layers of orchestration with agents running other agents in a very fixed scope with very specific instructions.
I give my agents personalities
Certain agents and AI outputs as a default can be either bland — or worse — they can be sycophantic. If I tell certain models that I’m planning to jump over the Grand Canyon on my bicycle, it might give me words of encouragement.
Some people have said that to work around that, you can tell it to be a blunt business advisor that acts rude if it needs to. So, I’ve tried agents with different personalities based on movie or TV characters.
My first test case with this was creating an agent that refers to me as “Captain” and believes we’re on the USS Enterprise from Star Trek. That was fun. Building it was easy with ClickUp’s Super Agent Builder.
I have an “Atomic Habits” agent that helps me stick to habits. I gave it the personality of the SNL fictional character and motivational speaker Matt Foley. I have an agent who helps me with vendor negotiations, and I gave him the personality of FBI hostage negotiator Chris Voss.
It doesn’t always work. For example, if you give an agent the personality of Yoda from Star Wars, it can lose its luster over time, and then you have to work with awkward phrasing in every single sentence. So we discontinued some iterations of different personalities. But I’ve found that it can be beneficial to model the agent on a real personality with knowledge that’s relevant to the task.
I try to make sure that anytime I’m going to give an agent a personality, it won’t interfere with any work that they need to do. So I’m not going to give any analytics-focused agents any personalities where they could possibly go rogue.
My work has fundamentally changed
Overall, my work has fundamentally changed and is continuing to change. A year ago, I’d spend a good amount of time pulling reports or building tables, and now I have agents doing all of that, and that are waiting for me when I sit down to start my day.
I spend most of my day orchestrating agents, reviewing their outputs, and giving further directions and when needed, like improving or iterating on prompts to help them get better outputs.
I’m spending less time on administrative work and more time on higher-level strategy work and reviewing agents’ outputs. I’m spending a good amount of my time on the orchestration of it all, managing a team that doesn’t sleep.
It’s a new kind of job, and I don’t think my role is disappearing. It’s just shape-shifting.
