Karan Kurani published the essay “Structure of Third Party Agentic Apps” about a new class of applications — third-party agentic apps (TPAA): third-party applications that run not as wrappers over an LLM, but inside an agentic harness — Claude Code, Codex, Pi, or OpenCode. In his scheme, the model plays the role of “hardware,” the harness — the “operating system,” and TPAA — the third classic layer, the application itself. The developer supplies only four components: dynamic skills, ready-made scripts, environment configuration with venv, and an empty database template; the agent writes everything else on the user's machine. Using the example of the local health application CareLoop, the author shows that the scheme already works, but its behavior depends on the “model + harness” pair.

What happened

Karan Kurani, author of the blog Experiments with AI, published the essay “Structure of Third Party Agentic Apps,” in which he introduces the term third-party agentic apps (TPAA) and describes their architecture. Such applications run inside an agentic harness — he names Claude Code, Codex, Pi, and OpenCode — and not as standalone wrappers over an LLM. The TPAA developer passes the agent only four artifacts: fully dynamic skills, ready-to-run scripts for efficiency, runtime environment configuration with venv, and an empty database template; the rest of the code, dependencies, and support are written by the harness itself. The illustrative case is CareLoop, the author's local health application: the agent, without his involvement, installed DICOM libraries to read the user's MRI scans and itself restores broken environments and databases after OS updates. The database structure intentionally diverges for each user, so instances of the same application stop being identical.

Context

The framework from the essay continues the classic three-layer computing scheme: the model performs the role of the processor, the harness — the operating system, and TPAA — the application layer on top of them. No new model capabilities are required for this: existing skills are used — code generation, tool use, working with the environment — the novelty lies in the architectural scheme of use. The product becomes not code with a fixed interface, but a specification according to which the agent assembles a working application for the user on their model provider. The author calls the working style meta-programming: support is reduced not to releasing releases, but to keeping the agent from going off the rails. At the same time, the trust model changes: security and privacy questions move from the developer's servers to the user's machine, which works with their own AI provider. There is no niche in the usual sense yet — without a TPAA store, format standards, and noticeable discussion.

Why this matters for the industry

If an application is skills, scripts, and a database template, then the software distribution platform becomes not the OS or the App Store, but the agentic harness, and compatibility with Claude Code, Codex, or Kimi Work begins to mean as much for such a product as compatibility with an operating system. The economics also change: the cost of development and support drops to almost zero, because the agent itself writes code, installs dependencies, and fixes environments — but the price is stochasticity, that is, runtime behavior that cannot be fixed by a release. Classic regression tests are poorly suited to this: instead of checking “does the code crash,” you need to check “does the agent go off the rails,” and the uniqueness of each user's database makes direct comparison of instances impossible. Hence the likely near-term shortage — methods for testing agentic behavior and metrics for “not going off the rails,” without which the TPAA pattern is difficult to apply anywhere beyond personal tools. If harness platforms begin to standardize the packaging of skills and environment configurations, the pattern may take shape into a niche of cheap vertical prototypes on local data without a server backend.

Why this matters for users

You can build your own application today and without programming: it is enough to describe the happy path in the form of skills, ready-made scripts, and an empty database template, and the harness will write the environment and code on the fly. The target audience of the recipe is those who already have an agentic harness installed and paid access to their own AI provider. A realistic starting product based on the CareLoop template is a personal tool on local data: receiving raw data like MRI scans, timelines and “question — next step” cycles, while the data does not leave your machine. The main pitfall is that the quality of work depends on the “model + harness” pair: in August 2026, the author observed that the Fable model in Claude Code sometimes refused health requests due to an internal “bio” category, so a refusal may mean not an application error, but a filter at the model level, limiting the scenario without the user noticing. Special caution is needed in medicine: CareLoop is an alpha version, and reproducing its recipe on your own medical data should be done with this in mind.

What is still unknown / limitations

The source only confirms the happy path of one custom application: CareLoop was created by the author himself, and there are no independent replications by third parties. This is an n=1 demonstration without measurements — the essay has no success rate, failure statistics, inference cost data, delays, or the share of blocked requests, so it is impossible to talk about readiness for load with SLA. The thesis of “four artifacts” should not be read as a working software distribution scheme: it is only confirmed that one author built and maintains one personal application this way. There is no resonance yet: the HN post got 1 point and 0 comments, there is no TPAA store or standards, so this is a reconnaissance signal, not a market. Forecasts at the level of “harness as a distribution platform on par with the OS,” trajectory evals, and compatibility conventions between harnesses are hypotheses to be verified, not established facts.

Sources

Author

Look at AI, editorial team