Our 2026 theme at Gigi is "Run to magic." When we say that, most people assume we mean building more flashy AI features. We are, and we'll keep doing more of it. But the less obvious part of running to magic is just as important: being intentional about what we're not building, and choosing reliability over the next shiny feature.

We pick a theme every year, like a chapter in a book. It tells us what we're optimizing for and, just as importantly, what we're saying no to.

"Running to hard" is a common concept in company building. It is the practice of prioritizing the hardest problems that matter, because solving them creates disproportionate value and lasting differentiation. The hard road is inherently challenging, but it also produces advantages that only exist on the other side. For that reason, companies are encouraged to run to hard.

Paul Graham, Founder of Y Combinator, suggests that companies should use "difficulty" as a guide. He clarifies what this means in practice:

"We deliberately sought hard problems. If there were two features we could add to our software, both equally valuable in proportion to their difficulty, we'd always take the harder one. Not just because it was more valuable, but because it was harder. We delighted in forcing bigger, slower competitors to follow us over difficult ground. Like guerrillas, startups prefer the difficult terrain of the mountains, where the troops of the central government can't follow. I can remember times when we were just exhausted after wrestling all day with some horrible technical problem. And I'd be delighted, because something that was hard for us would be impossible for our competitors." (link here)

These difficult problems that are exhausting to solve internally are often invisible to customers, yet they are precisely what makes products feel magical.

The best AI products feel like magic. We see this in our own product. When Gigi flawlessly executes a task for our customers, we create a magical experience. We palpably see the joy in our customers when Gigi saves them hours and delivers better customer outcomes.

In the same way that Paul Graham used "difficulty" as a guide, our ability to create magic for our customers is how we demonstrate value and generate alpha.

But magic doesn't always mean building the feature that shines in a cool demo. In fact, a lot of the hard work behind magic is in the paths we decide not to take.

No legacy tech

We have an order performance and full funnel dashboard in our product that we call our self-described legacy tech.

We included these in our product because we believed their current form was the bare minimum of legacy tech needed before we could begin building magic. We have not touched these two surface areas in our product since our launch last summer. Similarly, our customers cannot edit an existing campaign by pressing buttons. These are intentional design decisions that we've made, and will continue to make. More tradeoffs will arise and when they do, we should always ask: "Are we running to magic or are we investing in legacy tech?"

Making these decisions in a vacuum is easy. The pull of our customers has been alluring, and saying no when they ask for improvements to legacy tech surface areas is painful. Saying no to a customer who is asking for a real improvement is one of the hardest things to do as a founder. But we will keep doing it in favour of magic.

Being consistently reliable is magic

We've been able to demonstrate glimmers of magic by flawlessly executing tasks for our customers. Assign Gigi work, she will get it done quickly and reliably. But in instances when Gigi is unable to execute this work quickly and reliably, she provides a maddening experience antithetical to magic. Similar to a team member, if you assign someone work and they are unable to consistently complete that work at a high level, you will stop assigning them work. Every task that's assigned to Gigi is an opportunity to deliver magic and build trust. Every task executed suboptimally is a potential breach of that trust.

Over the last 6 months, we've relentlessly focused on Gigi's reliability, speed, and accuracy. Take latency as an example.

Latency has historically been a problem for us at Gigi. Response times for chat and ad hoc tasks when we first launched averaged over 2 minutes and could sometimes take 5-10 minutes depending on the complexity. We created a north star goal to hit an average response time of 30 seconds by April 1. We hit it. We made optimizations for tool calling efficiency, and the underlying models became more efficient. This was an ongoing effort by various members of our engineering team for 2 months, and will forever remain something that we manage.

Additionally, task failures are something that we religiously track at Gigi. Every week at all hands we look at the number of tasks that Gigi completes (note, each of the last 3 weeks we've had an all time high) and then we look at task failures as a % of the total tasks. The number of task failures is in the low single digits right now, but low single digits for us is unacceptable. Gigi needs to execute tasks flawlessly every single time. So, over the past month, we've set a vision zero internally for task failures, and for some of our team members we won't be taking on any new features in favour of achieving vision zero.

Latency and reliability won't win a pitch competition, and they won't deliver the "ah ha" moment in a customer demo. But like any team member, these are the foundational virtues of how an AI agent needs to work. When done flawlessly, it feels like magic. That's the chapter we're writing.

Note: This was originally published as part of our Cherry Picked newsletter. Subscribe here 🍒