An adage in ML has always been, *coffee stir*, “it is more art than science.” Today this is truer than ever. Building agents is a weird, non-determinstic process that adds a squirm-inducing level of variance into systems.

In these new waters, engineers would be well-served to hone their creative process. The best engineers will be artists.

1. Building isn’t linear. I’d bet that anyone who has spent time building agent infrastructure over the past 2 years will flashback to nights of tearing out their hair, cursing at an API. The internal conflict of “this is deeply cursed and will never work” to “obviously we’re going to figure this out” has been played out cyclically as foundational model improvements have rapidly streamed in. While still far from perfect, these reasoning systems are stable enough for integration into our tools and lives. And their promise, and an exponential takeoff that is already in place, is their proliferation into all of our tools as we enter an agentic era. This history of AI, machine learning, deep learning, etc has been that of non-linear development. 10 years ago as the commercialization of deep learning was taking off, engineers were made to quickly get comfortable with black box hidden layers, finding canaries in their outputs that flagged data issues, and stacking layers together like legos - all to throw darts at the holy dartboard of driving down losses *coffee stir*.

Today’s agent infra dev process is arguably even less linear. One has to somehow imagine how semantic meaning - given online data corpuses we can’t fathom - is translated to the black-box LLM - and what characters in the prompt should be manipulated to get sensical answers that pass evals most of the time. As LLM calls are chained, constant data checking must take place with ejects to call in user help if needed. And evals must be built up with careful user rollouts to try to preempt embarrassing agent failures, amplifying the paranoia of engineers good enough to know to be paranoid.

2. Building with a relationship to the work. With new manifestations of intelligence, come new types of relationships. As engineers build out agents that intelligently operate in new and unexpected ways, they must navigate a relationship to work that takes on its own relational qualities, like agent memory and understanding. What is it for your work to know users, and then you? What is it to build on a system that knows you? It’s been said that when people feel understood, they feel loved. And so what is it to love the systems we’re creating?

A non-linear process and a relationship to the work is the territory of the artist. Engineers, trained towards control, are compelled to swim further into uncertainty. We’re being asked to risk not knowing, to risk embarrassment, and to accept that our efforts might get sucked into a black hole that only produces spaghetti results.

We’re catching glimmers of playing creator to a system that understands. Humans sit in heart, mind, and body ways of being. While engineers have largely lived in a mind space, we're now called to a heart relationship with the systems we're building, a relationship to a creative spirit as the work lives.

A creative process buoys artists through the shit sandwiches they’re forced to stomach day in and day out to breathe life into the work. It looks like an entitlement to creating, a discipline to putting the time in, and an acceptance that one can only fully control the process, and not how it ends and lands. It’s a knowing that art is lifelong road, and the only relief is discipline in creating and the freedom in fully releasing ideas as they incarnate and leave.

Agent development pulls engineers closer to the headspace of artist - and my bet is if engineers embrace and protect their creative processes, we unlock the next level of innovation in society.