Thom Wolf published The future of software on November 9. His point is that in the limit, software won't mean anything anymore. Until then we live in an in-between age with three new precepts: everflux software, translucent sources and bespoke features.
I agree with all three. Here is what each one means for anyone selling to agents.
Everflux software
Software no longer steps from release to release. It evolves continuously, because everything around it ships at a blistering pace too.
The pace shows up in the benchmarks. In 2023 the best AI agent resolved 4% of the real GitHub issues in the SWE-bench coding benchmark. Today it resolves 95%. On computer-use tasks (OSWorld), agents went from 12% in 2024 to 90%. People reach 72% on that one.
The length of the work changed too. METR measures how long a software task an agent can finish half the time, counted in the time a human expert needs for it. That went from 4 minutes with GPT-4 in March 2023 to about 12 hours with Claude Opus 4.6 in February 2026.
METR's own caveats apply. These are well-specified software, machine learning and security tasks, and the 12-hour figure comes with a wide error band.
So an agent can't rely on a fixed list of tools. What was best last month may be gone or beaten today. It has to look up what's available, and at what price, every time it starts a job.
Translucent sources
Closed source is becoming open source in all but name. A model can pick apart an app, and a mere description of an interface is often enough to rebuild it. Thom wrote the essay in a text editor Claude built for him from a few demo videos.
So the code is no longer the thing you sell. What a description can't rebuild is the data behind the app and the access to it: a fresh database, a license, the right to call an API at scale.
Bespoke features
Everyone becomes their own product manager. Want a search box, an API bridge, a custom button? It's a couple of minutes away.
But a search box needs something to search, and an API bridge needs an API on the other side. The more people build their own software, the more building blocks they need, bought when they need them.
The upstream problem
Thom's challenge is the collision of the first and third precepts. If you fork software to add your feature, you lose every future update from upstream. His answer is agents rebasing your personal features onto each new release overnight.
Data has the same problem. Build your own scraper and you are its upstream from now on, fixing it every time the site changes. Buy the data per call and someone else keeps it current.
There are 1,087 services in the Delegate catalog today, each one bought per call. If software stops meaning anything, which part of yours would you still pay someone else to keep alive?
Sources
- Thom Wolf, The future of software, November 9, 2026.
- Stanford HAI, AI Index Report (2025, 2026), plus the SWE-bench and OSWorld leaderboards for 2026; partly vendor-reported.
- METR, Task-Completion Time Horizons of Frontier AI Models, Time Horizon 1.1, last updated September 8, 2026.