Как собрать AI-харнесс для SaaS-разработки: связка Orca, Pi Agent и GitHub

A developer published on Habr in July 2026 a detailed breakdown of his own AI system for SaaS development — a so-called agent harness. The main thesis: in real production work, not the most powerful language model wins, but the correctly built infrastructure around it, including the Orca orchestrator, lightweight Pi Agent, GitHub, and a custom VPS pipeline.

Why the harness matters more than model choice?

The debate over LLM selection never subsides: GPT-5 or Claude Opus 4.6, Gemini or Codex. But the author puts forward a radical thesis: in real development, the question "which model to choose" is secondary. Primary is the question "how is the model embedded in the workflow."

Agent harness is a layer between the developer and the language model that manages context, task routing, isolated code execution, logging, and feedback loops. Without it, even the most powerful model remains a smart chatbot, unable to reliably work in a production environment.

What components make up the setup

The author describes a specific architecture of five elements:

• Orca — central orchestrator: receives a task, breaks it down into subtasks, and routes to the appropriate agent or tool
• Pi Agent — lightweight execution agent for targeted tasks: code fixes, refactoring, logic verification
• GitHub — repository as the "single source of truth" for the system and synchronization point for agents
• VPS — isolated execution environment: agents run code on a remote server without risk to the local machine
• Custom pipeline — describes the task flow from assignment to merge

Orca and Pi Agent work as a pair: the orchestrator sees the big picture and distributes work, Pi Agent executes a specific, clearly defined instruction without needing to maintain the entire project context.

GitHub plays a dual role in this scheme. First — the familiar code repository. Second — synchronization point: agents write to branches, create pull requests, and the developer reviews changes in a classic code review loop, only now AI is on the execution side.

How this approach differs from regular ChatGPT

Most developers use LLM in "ask-answer" mode: paste code in chat, get a fix, copy back. This works for isolated tasks but breaks down in system development.

The harness changes the paradigm: the model becomes part of an automated workflow where each step is verified, versioned, and logged. An agent error is not lost — it's captured in the system and can be debugged. VPS isolation: multiple agents can work in parallel without conflicts on the local machine.

"In real work, it's not just the strongest model that wins, but a

correctly assembled system around it" — the author's key thesis.

What this means

The article captures an important shift in AI development practice: the growth point becomes not model quality, but infrastructure maturity around it. Developers who invest in harness — orchestration, isolated execution, versioning — get stable production results. Those who simply switch between ChatGPT and Claude work slower and less predictably.