Agentic Coding — Directing Agent Teams

Agentic coding is the practice of directing autonomous coding agents the way a tech lead directs a team: with clear specs, tight review loops and full accountability. It is my daily method for shipping software, and I consider it the single most important engineering skill I have built in the last three years.

My toolchain is deliberate. OpenCode for fast, model-agnostic pair programming in the terminal. Codex for parallel execution across isolated sandboxes. Claude Code for deep, codebase-wide planning and refactors. Hermes Agent — my own stack — routing models and orchestrating the other three. Different tools, different strengths, one workflow.

The core loop

0

Brief the agents

0

Generate & iterate

0

Human review

0

Deploy & monitor

The loop is exactly the four steps described in The Agentic Method: brief, generate, review, deploy. What changes per tool is where each step happens — OpenCode lives in my terminal next to vim, Codex fans work out into sandboxes, Claude Code plans before it touches a file, and Hermes coordinates long-running tasks and keeps everyone supplied with the right model.

What agents changed about my engineering

• Specs became a first-class artifact. A precise brief is worth more than a clever prompt.
• Review became a superpower. Reading diffs fast — and suspiciously — is now a core skill.
• Architecture matters more, not less. Agents amplify good structure and expose bad structure instantly.
• Testing got serious. If the test suite is weak, agents will happily produce confident garbage.

Proof, not promises

This portfolio — two locales, a cloned VitePress theme, custom SVG icon set and 14 generated illustrations — was built through this loop. So were NetPilot, OmniRouter and every panel in the projects section.

The following pages document each tool: OpenCode, Codex, Claude Code and Hermes Agent.