AI Agent Practical Notes: Engineering Practice from CLI Workflow to Local Governance

Introducing the open-source project ai-agent-field-notes, maintained by DevOps consultant Abdellah MOUHTAJ, focusing on practical experience in AI agent integration. The project's core philosophy is "Test unstable tools, stabilize workflows, record failure modes, publish reproducible runbooks", using the four-step methodology of test-fail-stabilize-record, covering CLI workflows, GitHub integration, local governance, etc., and emphasizing the auditability, reproducibility, and secure governance of AI agents.

Original author/maintainer: Abdellah MOUHTAJ; Source platform: GitHub; Original title: ai-agent-field-notes; Original link: https://github.com/Abdel-MOUHTAJ/ai-agent-field-notes; Release date: June 4, 2026. The project is positioned as a series of DevOps practical notes, differing from ordinary tutorials in three aspects: it is not a pile of code snippets but complete workflow records; not an idealized tutorial but real experience of pitfalls; not a leak of sensitive information but desensitized general practices. The goal is to prove that AI agents are auditable, reproducible, and governable, suitable for actual DevOps scenarios.

Each note follows a four-step process: 1. Test: Reproduce integration scenarios in real environments, focusing on details not mentioned in official documents; 2. Fail: Proactively find failure points (silent failures, misleading successes, gaps between documentation and reality); 3. Stabilize: Add mechanisms like wrapper layers and checkpoints to ensure clear system feedback or entry into a safe state under exceptions; 4. Record: Publish reproducible runbooks, including failure hypotheses and lessons learned, explaining "how to do it", "why", and "what if not done".

As of now, 4 notes have been published:

1. Note 001: Integration of Google Jules and Antigravity CLI, covering authentication mechanisms, GitHub configuration, Git atomic operations, and local governance;
2. Note 002: Jules returning results via Antigravity CLI, emphasizing explicit repository specification, read-only retrieval, and manual control of key write operations;
3. Note 003: Governor Memory (Governance Memory), involving standardized memory formats, sensitive information desensitization, and rollback-supported writing strategies;
4. Note 004: mvp-github-writer, focusing on SkillOpt optimization, controlled Markdown output, and local experimental environment setup.

The project emphasizes governance and security in AI agent integration:

1. Authentication chain management: Securely manage API keys, OAuth tokens, and SSH keys;
2. Human-machine collaboration boundary: Adhere to "human-in-the-loop verification" — AI generates suggestions but key operations retain manual approval;
3. Failure mode documentation: Systematically record failure hypotheses and accumulate "negative knowledge" to avoid repeating pitfalls.

The project uses the following technology stack: Google Jules (AI programming agent), Antigravity CLI (local tool), GitHub CLI (gh), Git (emphasizing porcelain commands to ensure atomicity). The toolchain selection focuses on scriptability and automation for easy integration into workflows.

The project has reference value for the following scenarios:

• Enterprise AI agent implementation: Practical reference for authentication, governance, and security boundaries;
• CLI tool development: Considerations for designing CLI tools that collaborate with AI;
• DevOps practices: Reference for Git workflows, CI/CD integration, etc.;
• Security compliance: Practical reference for human-in-the-loop verification, sensitive information processing, operation auditing, etc.

ai-agent-field-notes represents a pragmatic attitude towards AI agent applications: no show-off, solving real problems; no avoidance of failures, systematically recording experiences; no boasting about AI omnipotence, emphasizing human-machine collaboration and governance frameworks. In the era of rapid AI iteration, such engineering practice records are valuable practical guides for teams and individuals exploring AI agent implementation.