Why AI-Native Companies Are Deleting Software You're Still Paying For (The $56K Lesson)

AI-native companies stall less because they stop paying the “abstraction tax” that hides work inside opaque, click-driven workflows. Even with strong model memory, agents hit a wall when the organization’s operating environment is built on hidden state—draft modes, permission gates, CMS screens, ticketing portals, and tribal knowledge—where the agent can draft plausible text but can’t reliably advise, execute, or ship changes. The result is that buying agents often means buying a conversation layered on top of the same bottlenecks, not a path to real throughput.

A concrete case comes from Cursor. Lee Robinson migrated cursor.com from a headless CMS back to raw code and markdown. The move was expected to take weeks (and possibly an agency), yet it was completed over a weekend using roughly $260 in tokens and hundreds of agent-driven pull requests and calls. The headline isn’t that agents are magically fast at coding; it’s that the CMS introduced a new kind of friction. Once a CMS sits between the agent and the underlying work, updates become “clicky” UI navigation instead of direct delegation to agents. Agents thrive when the work is visible and editable; the CMS became a wall as agent capability increased.

The deeper lesson is economic: abstractions used to be a fair trade for non-technical teams, but they become expensive when agents need to operate on the workflow. A CMS hides files, structure, deployment details, version control, and the messy mechanics behind a friendly “edit/publish” screen. When agents enter the picture, that hidden machinery turns into costly dependencies: identity systems split across tools, permissions require extra wiring, previews become brittle, sharing “what’s about to ship” adds friction, and uptime/operational complexity grows. Cursor’s own CMS usage reportedly cost $56,000 since September, a markup for graphical convenience.

Robinson’s fix wasn’t anti-marketing or anti-UI. It was a shift back to legible primitives: artifacts that humans and agents can both see, review, undo, and validate. The site’s complexity collapsed into a single inspectable substrate where changes are reviewable via diffs and traceable through version history. Cursor’s advantage, then, isn’t only agent tooling; it’s a culture where technical fluency isn’t confined to an engineering department. Designers commit code, user management treats “designers as developers,” and internal workflows rely on shared primitives rather than handoffs.

The argument extends beyond websites. In the agentic era, “code-like” workflows win because software already provides the infrastructure agents need: version history, tests, rollbacks, audit trails, and clear system-of-record artifacts. Enterprises that keep work trapped in GUI state and human memory will get drafting assistants at best. The prescription is to teach “primitive fluency”—state, system of record, change records, checks, rollbacks, and traceability—so non-engineers can express work in agent-legible forms. That turns agent strategy from procurement into literacy: the winners won’t be the firms that buy agents, but the ones that can delete sacred, expensive workflows and safely move more of the organization into artifact-based execution.