Our AI Employee Won Our Hackathon and We're Not Sure How to Feel About It

TL;DR: We let an AI Employee participate in our company hackathon as a joke. It won. Now we run weekly AI-only hackathons because they're actually producing useful stuff with minimal effort on our part.

The Accidental Experiment

We're People's Grocers, a small agency building apps for startups. Recently, Karl Marx had a random thought: what if we included our AI employee in our internal hackathon? Not as a tool for the humans to use. As an actual participant.

This wasn't some grand vision about the future of work. It was literally: "Hey, we've got this AI employee we've been working on, let's see what happens if we let it compete."

Let me be clear: we're not giving it specific problem statements or success criteria. The brief is essentially: "You're participating in a hackathon at People's Grocers. Create something useful." Obviously the system has access to a great deal of context about our business. We gave it access to our internal docs, emails, and conversations, plus a dataset we'd curated from Reddit and HN about common engineering complaints.

Last month, this AI employee won our hackathon with a communcation tool for visualizing development roadmaps.

Why "AI Employee" and Not "Tool" or "Agent"?

You might be wondering why we keep saying "AI employee" rather than "agent" or "tool" or some other more typical term. There's a reason for this. As a contract development shop, our entire business model has an expiration date. It's not a speculative question of if AI will replace much of what we do—it's a matter of when. So for us, developing AI capabilities isn't some nice-to-have innovation project; it's a life-or-death priority.

Yes, the concrete implementation today is what most people would call an "agent" or a bundle of fuzzy logic workflows. But we deliberately call it an "AI employee" because it orients our thinking toward where we believe this is headed, not because we've already arrived. The roadmap from today's collection of special cases to tomorrow's truly capable system remains long and winding.

If you're in the business of selling labor (which we are) rather than owning stuff, then you should be at DEFCON 1 - working in shifts around the clock. An AI employee is the last stop of ownership train are you getting on?

So How Did It Win?

The AI came up with a communication tool—specifically, a new type of diagram for communicating roadmaps between engineers and stakeholders.

It's not a software program. It's an algorithm for structuring your communication. Just like a template for conference talk slides. The capability mapping system plots features on two axes (user visibility vs. engineering robustness) and shows connections between different capabilities.

This made it immediately obvious when we were over-engineering features or when we had gaps in our plans. It also visualized uncertainty in implementation paths rather than hiding it.

We've published a detailed breakdown of the project here. The remarkable thing is that the AI's output was essentially complete—we didn't have to do anything significant to it. The blog post explaining the system was also written by the AI (with only light redaction of sensitive internal details by a human). You can judge the quality of the work for yourself.

Weekly AI Hackathons

We now run AI-only hackathons every week. After that first win, it just made sense. Aplogies in advance for stating the obvious, but we're pretty excited about this. Of course the AI system only takes around 30 minutes to produce the ~1,000 lines of code you might find in a typical project. These days the judging is the longest part - we spend an hour or two reviewing the different solutions.

We're sticking with the weekly schedule while we continue improving the AI employee system. It gives us a good way to track progress and see how our changes affect what it produces.

Why a Communication Tool Instead of Code?

The most interesting aspect of our AI employee's winning submission wasn't just that it created a useful tool - it's that it identified and solved a real communication problem rather than defaulting to writing more code.

Our AI employee correctly picked up on the fact that People's Grocers puts significant emphasis on communication precision. While we write plenty of code, the niche we've carved out is translating complex technical capabilities into intuitive explanations for clients.

What's remarkable is that we didn't specifically train the AI to focus on communication problems. The models are practically stock. We did the normal task-specific fine-tuning on a couple thousand examples that everyone is doing these days - nothing special. Looking at the thinking tokens and some mechanistic interpretability features hints of how generalized the concept of a "problem" is for the base model.

The AI employee has access to several internal resources - notably our curated dataset of annotated engineering blogs, several communication tool prototypes, and training. Maybe it "saw" that our clients typically come to us with technically sophisticated products they're struggling to make intuitive. As always, we only need the one rollout which successfully generalized from this information to identify a gap in our toolset.

This demonstrates something powerful about current reasoning models: they can extract the underlying patterns of how a business operates and identify non-obvious problems worth solving. The AI didn't just automate an existing process or write more code - it invented a new methodology for communication. That's what's exciting to us about this experiment. The AI wasn't constrained by the assumption that a hackathon solution needs to be code. It recognized that sometimes the highest-value contribution is a new way of thinking about and structuring information.

Why Aren't More People Doing This?

Honestly, we're not sure. Maybe it sounds too gimmicky? Maybe people are worried about the optics of "replacing humans" in creative work? Or maybe everyone is doing it and just not talking about it?

Are other companies doing similar AI-only hackathons? Let us know at hackathons@peoplesgrocers.com. We're curious if this is just a weird thing we're doing or if there's something actually interesting here.