Agentic coding update · AI slop is real, the self-improvement loop, and our cleanup agent

99.9% of our codebase is written by Codex and Claude Code.

We spend $1.2k in tokens / month.

Before the job-fearing devs spam the comments with objections, hear me out.

AI slop is still very much real.

But software engineering with the right instructions, guardrails, skills and MCPs is more or less solved.

Especially in our case. OXYGEN is a CLI and MCP for GTM. 95% of the actions are actually verifiable and the agent can test them directly.

So here is our semi-functioning self-improving loop:

→ Agent builds the feature → Auto-deploys → Uses the CLI to test the live deployment → Finds the bugs → Patches and redeploys

What's NOT working:

Architecture still over-bloats. Massively.

→ Files balloon. 800-line components when 200 would do. → Brownfield context bloat is real. The agent walks into existing code, can't hold the full picture, and spawns 3 new components instead of editing the right one. → Over-engineering by default.

Our fix: a dedicated cleanup agent. Single purpose. No new features. Just walks every feature we've shipped and simplifies it. So the next coding agent that touches that area starts from a cleaner base.

We slowed down a bit. Worth it.

But still: hitting my limits faster than ever. Maxing out 2 Codex subs at the same time.

Are you team Claude Code or Codex? Why?