🛠 Agent writes the app instead of the developer: a new class of software
Karan Kurani published an essay, “Structure of Third Party Agentic Apps,” about third-party agentic apps (TPAAs) — third-party apps that do not work as wrappers over an LLM, but inside an agentic harness: Claude Code, Codex, Pi, OpenCode. The model plays the role of “hardware,” the harness — the “operating system,” and the TPAA — the familiar third layer of “applications.” The developer supplies only four components: dynamic skills, ready-made scripts for efficiency, environment configuration with venv, and an empty database template — the agent writes the rest itself. CareLoop, a local health app, is built on this scheme: the agent, without the author’s involvement, installed DICOM libraries itself to read the user’s MRI scans, and it itself fixes environments and databases that break after OS updates. Each user’s database is intentionally divergent in structure and incompatible with others, the app works with the user’s own AI provider, so the author has zero support cost.
🌍 If an app is skills, scripts, and a database template, the software distribution platform becomes not the OS or the App Store, but the agentic harness: compatibility with Claude Code, Codex, and Kimi Work starts to matter more than compatibility with an operating system. Testing shifts from the question “does the code crash” to “does the agent stay on track” — the author calls this meta-programming. The viability of such apps also depends on one-sided platform decisions: in August 2026, the Fable model in Claude Code sometimes refused to execute health requests, classifying them into its dangerous “bio” category.
👤 Today you can already “assemble” an app for yourself without programming: describe the happy path as skills, scripts, and an empty DB template, and the harness will write the environment and code on the fly. CareLoop shows the recipe — a local health app on your own data: raw data intake, timelines, “question — next step” cycles. The catch: the result depends on the “model + harness” pair, some requests may be blocked by filters, and so far only the happy path of one app has been confirmed — this is an experiment, not a ready-made software distribution standard.
Source 1: https://karankurani.com/writing/post/w_20260722063839_af741a/experiments-with-ai-structure-of-third-party-agentic-apps/ Source 2: https://karankurani.com/writing/post/w_20260721140414_b0e102/the-body-remembers-but-the-doctor-cannot-experiments-with-ai-in-personal-health/