MyRealTrip Builds AI Agent-Centric Development Environment… Shares Harness Engineering Case Study

Over the past year, MyRealTrip logged approximately 1,700 commits across 7 repositories as it charted its AI coding journey. Initially, AI was used as a supplementary tool, but now the company has applied the Harness Engineering concept, shifting to a structure where AI agents take the lead in writing code while humans handle direction-setting and verification. In February 2026, the order service repository saw 291 commits in a single month, with a significant portion generated by AI agents.

Four Stages of Evolution in Development Methodology

The development process can be divided into four major stages. In the first half of 2025, it was the Copy-Paste AI Coding stage, where code was generated from prompts and then manually revised by humans. In the second half of 2025, the process moved into the Augmented Coding stage, where Claude-based automated reviews and test generation were introduced. In Q4 2025, Spec-Driven Development (SDD) became established, with writing spec documents followed by AI generating code and tests becoming the mainstream approach. Starting Q1 2026, the company entered the Harness Engineering stage, building a system where agents manage documentation, code, and tests in their entirety.

A Map for Agents: docs-tree-tools

To enable agents to work effectively, the company independently developed a Claude Code plugin called docs-tree-tools. This tool focuses on structuring repositories in an agent-friendly way. Its key features are as follows.

• Automatic docs/ Tree Structuring: Uses index.yml as a central registry to map document IDs to paths, allowing agents to automatically load only the documents needed for their workflow.
• Doctor Diagnostic Engine: Scores index consistency, guide hygiene, Harness principle alignment, and link quality to manage documentation quality. It issues warnings when external references (Notion, Google Docs, etc.) are detected.
• Reverse Code Analysis: Automatically extracts glossaries, use cases, specs, and more from source code to keep documentation and code in sync.

Key Lessons and Outlook

MyRealTrip emphasizes that instead of relying on massive prompts, it's important to make the knowledge within the repository the Single Source of Truth. Information outside the repo is effectively nonexistent to the agent, so all context must be stored in a version-controlled form. Additionally, to ensure the reliability of AI-generated code, the company requires user story-based E2E test suites as a must. Going forward, the core of development will be less about the ability to write code and more about the ability to design environments in which agents can work, while design decisions and business logic verification will remain the domain of humans.