Jan 4, 2026 · 4:42 PM UTC
Great read. Yes, I don't think the models will subsume the application layer or the environments. Not this year, not ever. There's actually no intelligent system without a compatible environment. We are going to spend a few more years building agents.
I also see a need for vertical task models (like labellers, routers, agentic tool callers). Every company building agents has 100k+ traces coming in daily, and nobody is doing anything valuable with them. My guess is that Meta bought Manus not for the product they bill but for the traces. They now have the data on how millions of users actually directed agents to use tools and to perform tasks. (can share more...)
Also as labs explore uncharted territories like WFM, I see high quality data annotations, preparing datasets with information dense real world physical data, automating the training pipelines, moe routers all becoming important. There will be a lot of vertical developments this year.
really good read!
"No one tells you that you were living in the good old days until they are gone"
amen!
Love the point on models collapsing all customer-facing roles! We’re seeing this in GTM via account-specific agents goal-directed on the next step in the customer journey, not arbitrary functional boundaries (marketing, SDR, AE, CSM).
Great article! I just put out a piece with some related thoughts: nitter.cf/zivreichert/status/200…
Yep tech is no longer the main bottleneck issue - it’s ingenuity, talent and agency to execute better.
Great read. Am I the only one who's stuck on the map having the Golden Gate Bridge in the wrong place? ;)
great read! re:wabi, what do you think is their moat and companies like them? if nikita's point is true, does that not invalidate companies trying to make it fun to build things?
Wabi is a network effects product - creation and consumption all happen on one platform.
On the latter point I think its more of an awareness gap than a desire gap - even if the ratio of software creators remains power law there will be an enormous market of people who want to consume what they make.
I agree the bottleneck is shifting from “how do I build” to “what do I build,” but I think there’s a deeper failure mode: legitimacy.
When everything is cheap to execute, the hard problem becomes who gets to decide what’s worth building and most enterprises aren’t designed for that level of option density.
Do you think Anthropic Artifacts and OpenAI canvas are technically thin “exploration” apps, that could become thick in a similar way as Notion became for Google Docs? Non-coding use cases that can likely be extended beyond “word processing”.
Excellent post! There’s a fascinating question about experience too. Tomorrow’s apps (however narrow) will be built by those experiencing the pain points and then solving for them without being tethered to engineering “nerds”. How does the experience economy change as a result?
Right now almost all of our “tools for thought” are really tools for production:
IDEs for code, Figma for artifacts, spreadsheets for models. Even LLM chats are mostly glorified brainstorm machines.
But once coding agents get good enough, the real question is not “can we ship this?” but:
Should we ship this, for this customer, in this market, with these risks, right now?
That’s a different class of product than today’s copilots.
A real thinking tool would:
- Produce a small set of options, not a firehose of ideas.
- Attach constraints: resources, risk, dependencies, regulatory landmines.
- Attach evidence: telemetry, tickets, revenue, cohort data, experiment results.
- Force falsifiable predictions: what we believe will change, by when, and how we’ll know.
Until tools do that, “exploration” is just aesthetically pleasing brainstorming. The bar for “thinking tools” is much higher than “I had a good chat with my AI.”
Awesome read! Regarding 'Thinking Tools' - alot of development has happened on dev tools front, but the non - technical side remains a blue ocean - for example building for non-technical founders, creators, teachers, architects etc
For folks wanting to go deeper into building 'Thinking Apps' - do check out the paper 'Augmenting Human Intellect' by Doug Engelbart of Stanford Research.