I tried Vibe coding an AI coding harness.

Here's exactly how it went...

I needed a harness with:

• A TUI

• An agent loop

• Tools

• MCP support

• Skills

• Slash commands

• Custom rendering

—

So I started by dumping the whole spec into Claude Code and let it generate everything in one shot.

But:

• Slash commands weren't rendering properly

• The Skill command was misbehaving

• Features became hard to extend

• TUI flicker

—

In other words, it failed.

I figured you couldn't one-shot entire apps.

But there's a catch...

You CAN one-shot big features if you slice them correctly and run them through the engineering process.

—

So I built Squid.

An opinionated six-agent Claude Code setup designed to ship features like a real software team.

Here's the setup:

1. PM → Translates the feature documentation into atomic tasks

2. SWE → Uses TDD to write code

3. Tester → Runs verification and adversarial edge-case testing.

4. PR Reviewer → Reviews the diff.

5. On-call → Loops on CI/CD until it passes.

6. Self-improve → Scans the run for lessons and proposes updates to claude.md and memory.

—

Here's the gist:

The difference between vibe coding and agentic coding lies in the model's lifecycle.

• Human gates.

• Retry caps.

• Review loops.

• Architectural memory.

• Shared vocabulary.

That’s the system.

—

P.S. I break down the full Squid setup in this week’s Decoding AI Magazine.

It goes live tomorrow (12th May) at 1 PM CET.

Subscribe to get it delivered to your inbox: decodingai.com/p/squid-…