Jack Flintoft

The graph above shows the tectonic shift happening right now. Context windows (basically how much “input” data a model can ingest at once) have 10x-ed in the past year. That means a model like Llama 4 could, in theory, read Capital Volumes I and II in one go. Not quite Volume III (sorry Marx), but we’re getting close…

This is game-changing for enterprise software.

Before, automating workflows required structured inputs, brittle APIs, and rigid schemas.

Think of it like this: if a startup wanted to build a tool to help with customer support, it had to rely on narrow, structured inputs – maybe just the last few messages in a ticket, or the raw form data. The model couldn’t see the broader arc: the full conversation history, the internal back-and-forth, the subtle patterns of product friction, or how one issue echoed another from weeks ago.

It could observe a day of work. But never the shape of a workflow, let alone a job.

That meant it couldn’t understand what made a task hard, repetitive, or brittle. And because it lacked context, it failed at the edges – the one-off problem, the unusual escalation, the thing that makes up half of real work. Tools broke down where human labor actually lives.

Even when a model sat inside the system, it remained blind to everything around it.

That was mostly due to two limits:

• Agents weren’t good enough to generalize across messy, unstructured tasks.
• Models couldn’t hold enough context to make sense of the full environment.

Now, both of these have changed.

Agents are improving fast. And large context windows mean tools can passively watch full workflows – mouse movements, screen content, conversations, form fills – and learn directly from the ambient data.

This can unlock a new kind of wedge: A startup can enter a crowded market with one tightly scoped feature, train on what it sees, and rapidly expand horizontally – building agents around the full stack of work.

Two quick examples for clarity:

1. HR onboarding
   • A tool starts by verifying employee documents. While running, it sees how HR reps write emails, flag issues, update databases. With enough context, it begins surfacing suggestions. Then pre-fills forms. Then automates entire onboarding flows.
     
2. Patient intake in clinics
   • Epic dominates this space (listen to Acquired!). It’s tough to break in. But say a startup wedges in with a scribing tool for behavioral health clinics. Once inside, it watches clinicians move across tabs, talk to patients, chart notes. That data becomes training fuel. Suddenly, the tool can build agents to handle billing, scheduling, follow-ups — they build from the ground up.

In both cases (yes, they’re administrative), the initial feature is the wedge, but the ambient data becomes the moat.

Startups that understand this shift won’t just integrate into workflows – they’ll learn from them.

Context is everything.

---

Inspired by Jesse!