Experiments with AI – Structure of Third Party Agentic Apps
Introduces "Third Party Agentic App (TPAA)" as a new app paradigm where developers ship expected output shapes and behaviors rather than hardcoded functionality, with the agent harness executing the rest TPAA consists of four core components: dynamic skills, scripts, execution environment configuration, and a database — all modifiable by the AI and customizable per user The architecture mirrors a three-layer computing model: frontier LLM model (hardware), agent harness (OS), and TPAA (third-part
Analysis
TL;DR
- Introduces "Third Party Agentic App (TPAA)" as a new app paradigm where developers ship expected output shapes and behaviors rather than hardcoded functionality, with the agent harness executing the rest
- TPAA consists of four core components: dynamic skills, scripts, execution environment configuration, and a database — all modifiable by the AI and customizable per user
- The architecture mirrors a three-layer computing model: frontier LLM model (hardware), agent harness (OS), and TPAA (third-party app)
- TPAAs are harness-agnostic but harness-dependent — they run inside agent harnesses like Claude Code, Codex, or Kimi Work, but cannot function with raw LLM models alone
- A key innovation is self-healing capability: the app automatically updates, fixes broken databases, adapts to OS changes, and resolves configuration conflicts without developer intervention
Why It Matters
This article proposes a fundamental shift in how applications are designed and distributed — moving from shipping executable code to shipping behavioral specifications that an AI harness interprets and executes. For AI practitioners, this redefines the developer's role from writing detailed logic to architecting data shapes, guardrails, and skill definitions. It also highlights the emerging fragmentation risk where app quality varies significantly across different model/harness combinations, which has direct implications for UX consistency and developer strategy.
Technical Details
- TPAA Architecture: A self-contained bundle comprising skills (modular capabilities that interact dynamically), pre-shipped scripts for efficient execution, a virtual environment (venv) for secure isolated code execution, and a dynamic database that diverges per user over time
- Four Required Components: Dynamic skills, scripts, execution environment configuration, and database — all four are essential and removable without breaking the app; each is modifiable by the AI and scoped to the user's use case
- Harness Compatibility: Works across multiple agent harnesses (Claude Code, Codex, Pi, OpenCode, Kimi Work) but output quality varies substantially between them — e.g., Claude Code's safety filters may refuse health-related queries classified under "dangerous bio" categories
- Self-Healing Mechanism: The app automatically detects and repairs failures caused by OS upgrades, conflicting configurations, or database corruption by modifying its own code, refreshing databases, and adjusting installed packages on the fly
- Database Design: Ships as an empty template with usage instructions; the actual schema evolves dynamically through agent-driven interactions, making each user's database unique and intentionally incompatible with others
Industry Insight
- The TPAA model could dramatically lower the barrier to app development, enabling non-technical users to benefit from customized AI applications without requiring traditional software engineering — but it also shifts quality control from developers to the underlying model/harness combination
- Developers should expect a fragmented user experience landscape where the same TPAA performs differently across harnesses; strategic model/harness selection will become a critical design consideration, and benchmarking across harnesses will be essential
- The self-healing paradigm reduces long-term maintenance burden but introduces new risks around autonomous code modification and database divergence — robust guardrails and rollback mechanisms will be necessary before TPAAs can be trusted in production or regulated domains like healthcare
Disclaimer: The above content is generated by AI and is for reference only.