February 13, 2026
When I think about the impact of AI on human work, the metaphor I usually point to is a ladder. The lowest rungs look like brick laying. It's making the thing. The higher rungs of the ladder look more like strategizing. It's figuring out what to make.
One of the magical effects of AI is that it lifts us up the ladder. The engineer who used to write lines of code increasingly works at the level of defining specs. The work rises from coding to planning.
For engineers, this has meant a euphoric sense of productivity. Devs are talking about 10x code output. Some notable teams are even reporting that they've stopped using IDEs altogether.
There's a dark side to all of this. While engineers are currently at a happy rung of the ladder, what happens at future stages of AI progress? We have to consider that agents might not stay at the lowest rungs–that they might climb the ladder too.
Recently, a friend of mine was told at their onboarding for a large frontier lab: "Welcome to the last two years of your career."
I earnestly love software startups. The dream of starting and investing in important software businesses has animated my professional career for the past 15 years. It's not easy for me to be objective about this. I really do not want to believe that the game is going to be over soon but I also know that my feelings don't matter. The GPUs will continue to hum. The models just want to learn. The agents will continue to climb the ladder.
What then is left for workers, the tools they use, and the entrepreneurs and investors who profit from the business of tool building?
My general thinking is that when agents automate work, applications can follow users up the ladder or transition down to infrastructure.
Following users up the ladder is what Cursor has done. Cursor started as a copilot, transitioned into multi-file edits, autonomous tasks, and now they're involved in agent orchestration… A bullish view on Cursor entails that even if agents take on higher-order levels of work, the company can successfully meet the human user at the level of managing the creation of software, even to the terminal point that the factory is the product and humans are replaced or reduced mostly to sensors (we'll know things that the models won't).
One way to invest into the face of relentless progress is the optimism that there will still be value to offer at the highest possible rung of the ladder. Borne by this kind of optimism, an entrepreneur has to commit themselves to building the primary interface where a company defines their strategy for a particular domain of work. The phrase I once used to describe this was that strategic leverage accrues to the place where work is defined. Recent progress has forced me to adjust my language.
Another way is to build infrastructure. You accept the bitter forecast of human work for your domain and transition to a customer population with a longer shelf-life: agents. If you're not in the value chain of token production (e.g. frontier labs, data centers, RL environments, data brokers etc…), the way to do this is to help agents solve problems in a way that is comparatively token-efficient or suits agent ergonomics.
The last way is to build agents yourself and hope that your business can compete against the biggest labs–bear in mind of course that these companies, with all of their capex commitments, have an economic imperative to capture as much value as they can from the agent layer, even if it means totally cannibalizing their API customers.
The main result of automating software tasks is that people can build tools relatively cheaply. The cost reduction comes from tokens being cheaper than human labor.
Does software being cheaper change competitive dynamics? The natural instinct is to say yes but that flies against most of what we understand about software competition.
The code was (almost) never the moat. If you asked the head of product at Snapchat if they're worried about competitors copying their features, the answer has to be yes, but that this isn't anything novel. Facebook has been copying them for the better part of the last decade. Cheap code isn't going to change this.
In the enterprise world, the situation is more nuanced. We should expect that companies are more interested in building their own tools than they have been in the past.
I remember talking to a friend of mine several years ago who was considering building an internal tool that looked a lot like a CRM. At the time, the idea was obviously misguided since it represented an unreasonable expense of engineering resources. But we're fast arriving at a world where every build vs buy calculus has to be recalculated.
If token costs are low enough that the cost of creation (and agent-powered maintenance) is less than the lifetime cost of consumption, then there's an obvious argument in favor of the DIY approach.
There is an important exception to this and I think this represents a critical trait that ultimately makes some kinds of 3p software durable against the threat of DIY (or even open source) alternatives.
Does the company possess scale advantages such that they can produce more useful tokens that improve the quality of their product than their competitors? And crucially, does the company operate in a domain where marginal increases in intelligence matters? I do not believe that this quality exists in most incumbent software categories.
I ask myself a lot these days whether software is still worth investing in or whether the labs are running out the clock on the game… My present view (self-serving as it is) is that agents are going to climb up the ladder in almost every vertical of digital work, and that when the music stops, the great companies will be those that represent the largest pools of compute dedicated towards solving the most important problems.
The future of the software firm is bright even when the sun sets on human work.