Two ways of thinking about agents are forming, and they lead to very different futures.
The first sees an agent as an extension of your own mind. Used well, it works like your system one: you decide on a goal and it gets carried out, the way your hand picks up a glass without you thinking about each finger. Follow that far enough, through brain-computer interfaces, and the chat box disappears. There’s no “other” to talk to, just less distance between thought and action. The catch is that an extension amplifies everything you bring, including the blind spots you can’t see.
The second sees an agent as its own entity: an independent being that makes decisions and works alongside you. In that world, chat might be the final interface, because talking is how we already work with colleagues, friends, and family. Voice or text, it’s the same thing: you talk to it the way you’d talk to a teammate.
Both views agree on what matters today: someone still owns the judgment. Whether the agent is part of you or next to you, the output is only as good as the person deciding what good looks like.
Someone has to hold the pillars
Every serious system has a few things that must not move: brand pillars, an architecture, a compliance rule. Agents are great at everything around those pillars and bad at respecting them over time. You make dozens of small decisions a day with an agent, miss one, and the drift compounds. Technical debt has stopped being technical. Any system run with agents can rot this way.
This is the whole case for forward-deployed engineers. Someone has to sit inside the customer’s business, know which three things can’t move, and keep catching the drift. Models will keep getting better. Until they can hold the pillars on their own, the human in the loop is the product.
There is no middle anymore
Hand an agent to an average performer and you don’t get average output. You get slop. The agent multiplies the number of decisions they make in a day, fatigue sets in, and they stop judging. Hand the same agent to someone with judgment and stamina and you get work that used to need a team.
The “good enough” price band is disappearing. There’s no longer a $30-an-hour middle that reliably produces adequate results; output is either slop or excellent. For me this changes what matters in a hire: how many good decisions someone can make before they get tired, and whether they notice when they’ve stopped.
Pivoting is a skill now
Startups spent a decade telling themselves pivoting is fine. It still feels like failure, which is exactly why we had to keep saying it. With models taking a real step every few months, that feeling is expensive. A strategy built for last quarter’s model can be wrong today, and the default is to notice six months late.
What I take from this: treat every major model release as a scheduled strategy review. Ask what got cheaper, what became possible, and which parts of the plan just died. Build around invariants that will hold for longer than three months, and let everything else bend.
What is a forward-deployed engineer?
A forward-deployed engineer (FDE) is an engineer who works inside the customer’s business instead of only on a product. They learn how the customer actually operates, build the solution with them, and stay to keep it working.
The role isn’t new. It’s the 90s hacker-consultant coming back: someone who could walk into an unfamiliar industry and tech stack and be useful within weeks. Agents just gave it a new name and a lot more leverage.
A good FDE is half engineer, half diplomat. Hackers call it social engineering; corporate calls it stakeholder alignment. Either way, an FDE who can’t get a room to agree loses most of their edge. At eigi, our forward-deployed engineers set up AI teammates inside the tools a company already uses and stay until the setup sticks, and I lead product alongside those pods.
AI-native means winning, not cutting costs
Most companies used AI to run the same support and sales cheaper, and customers noticed: the experience got worse. Being AI-native means using AI so customers like you more and competitors can’t keep up, like answering an RFP in two days while everyone else takes two weeks.
The other trap is cognitive load. Put a powerful agent in front of a typical employee and you’ve added decisions to their day, which feels like using a bad app. A few people get energized by that; most get anxious. So the hard part of AI adoption is design and change management. It’s why I like that our Eigis ask before anything goes out in your name: fewer decisions for you, and you keep the final say.
Learn AI from the AI
Paying for an AI course is the old way of learning something new, applied to the one subject where it makes the least sense. The model already knows itself. Ask it to build you a curriculum, tell it which parts bore you, skip ahead, and make it teach you at your pace. A recorded course can’t adapt to you like that.
Living in the foothills of the singularity
At eigi we think of the singularity as the gap between an idea and making it happen getting smaller every day. The side effect is that perfectly sensible decisions get invalidated fast. A plan that made sense in January can be wrong by April without anyone having made a mistake.
There’s an image I keep coming back to: rigid systems are rock, and the singularity shatters rock; the ones that survive behave like water. That’s the case for everything on my values page: experiment, move fast, care about the process, and expect to change your mind. I could be wrong about half of this page in six months. When I am, I’ll update it, with dates.