Edition 9

If you haven't been turned on to Tidemark's VSKP (Vertical Software Knowledge Project) by now, consider this your notice.

This article is about all the narrative on the death of the application layer by foundation models.

Tidemark's David Yuan argues the opposite: "AI is magic, but it needs an application layer to bring it to life for customers," and outside of coding, that layer largely doesn't exist today.

Why won't the foundation models build it in most markets? They concentrate where AI already works well, pre-training data exists, outcomes are verifiable, and a wrong answer is cheap. Coding checked all these boxes, so the labs built it.

Dental practices, accounting firms, freight brokerages, and title companies do not look like this, however, and for much of the task mix the right tool is still deterministic software, classic ML, plain rules, and other unglamorous real-world steps. Said another way, for the jobs to be done, which are just bundles of tasks, the model can only do part of it.

This matters because customers aren't paying for the AI parts, they're paying to have the whole job done. If you can't do the entire task mix, you can't 'do the work', hence, the need for the application layer.

Now, to be clear, Tidemark is not saying AI and agents don't work. Tidemark itself has transformed its own operations, but only after enduring the grunt work of enabling its tools and data, codifying its "semantics, taxonomies, and processes," and building a harness for predictable outputs.

Few (no?) HVAC shops or 12-person accounting firms are going to build and police an ontology of its own business: "it's a pain in the ass and it is nobody's job." That gap is the opening for the application layer.

Tidemark believes incumbents have more right to own the application layer as anyone, but cautions this should not be read as safety. It should be read as a window.

"Somebody is going to become the application layer — the System of Action — in your market. If it isn't you, you've got problems."

There will still be APIs etc., and underlying data structures that are deterministic, but software companies will have to build the AI harness and full solution for their customers or be subsumed by it.

With software architected around agents' operating workflows, intent = context + policy and reasoning = tools + deterministic state changes, adds Todd Saunders.

That is, "the AI winners won't just have the most data... they'll be the best at turning that data into reliable action. That is what an AI harness actually is."

The next era of software is one of applications rebuilt around the assumption that software itself is now one of the users; not just a chatbot on top of CRUD software.

ServiceTitan gave about 1,000 of its customers 30 days' notice that it is shutting off their integrations with Podium, a lead gen and management platform, in the middle of home services' busiest season.

Gokul Rajaram calls it "exhibit #1 in the fight that will define vertical software for the next decade: systems of record versus the AI-native companies building on top of them.", especially as migrations become quicker/easier, reducing switching costs.

An AI-native playbook has become enter through the integration, read the data out, then start doing the work the system of record only ever recorded (booking the job, answering the customer, chasing the estimate) until "the old system becomes an expensive database." The new era is making partner API terms increasingly strategic, as the incumbent sees the activity in their API logs and may decide to either rebuild their architecture around agents (a rewrite) or revoke access ("a Tuesday").

"An integration is not a product feature. It is a revocable treaty between two companies whose incentives can (and will) change." It's about incentives, not APIs, and for now, the system of record still owns the veto.

With some of the points made above on software being rebuilt around agents as a user, these fights over who becomes the application layer, not just who holds the database, will only continue. Everyone wants to be the "System of Action" (above) or the "ERX" (below) on the vendor side. Complicating it further is that, on the user side, most users want to bring their own agent, not yours (below).

Digital technology creates endless lightweight actions that feel productive but "produce weeds in your project garden".

Using Albert Einstein's prolific 1912-1915 push for general relativity as an example, Cal posits that meaningful accomplishment requires obsessively focusing on a very small number of projects.

Newport's fix is a "productivity purge": list your projects by life domain (professional, community, personal), star one or two per domain, and purge or dispatch the rest, then spend a focused month obsessing only on the starred few until they are actually done.

For ongoing/recurring obligations you can't just purge, create "autopilot" programming so they don't encroach on star-project time and to contain their "cognitive cost". Define when, where, and how you handle (e.g., weekly reports every Sunday at 5 pm).

Farnam Street's Outliers episode traces how Rockefeller went from a kid in a poor family in Cleveland, to knocking on doors for his first job and getting one in bookkeeping because of his good handwriting, to becoming the richest man in the world.