Why I run four coding agents at once

I currently work with four different coding agents in parallel. Claude Code, Codex, Antigravity and OpenCode. Not because I cannot decide. Because no single one is enough.

It feels like a new feature ships every week. The tools move so fast that the roles keep shifting. What was the best approach two months ago is outdated today. Use only one tool and you miss that.

The whole thing only works because they all reach into the same knowledge base: my Obsidian vault. That is where I document skills, workflows and project context. Local skills in Claude Code are gone the moment I work in a cloud session. So everything lives in the vault. Whichever agent I open, the context is already there.

Claude Code does the main work

The setup for this blog came entirely from Claude Code. Chatbot, embedding pipeline, edge functions, chat widget. Same for my career pipeline web app: 207 active jobs, Claude compares my profile against postings automatically and scores them. Or the landing pages I build for local businesses: scrape Google Maps, qualify leads, Claude generates a finished page in ten minutes. Headless CMS on top, deploy, done.

All Claude Code. But it has a blind spot: it rarely notices when it got something wrong itself.

Codex as the auditor

That is exactly what I use Codex for. Triangulation between models. I have Codex audit the output of Claude Code. Codex finds errors Claude misses. And the other way round.

A single model checks itself badly. Two models checking each other are considerably more reliable. I notice it every time Codex finds a bug Claude read past three times.

Antigravity for research

Since the 2.0 update I use Antigravity mainly for deep research, together with Perplexity. When I want to properly understand a topic before I build. Sometimes I write code with it too. Pragmatically it is also a question of tokens: on the Claude Pro plan I hit the limit regularly. A second strong agent helps there.

OpenCode for the routine work

OpenCode has no strong models hooked up on my machine. That is deliberate. I use it for work I do not want to spend my good tokens on: Obsidian cleanup, checking transcripts, small scripts. For my bachelor thesis I built my own audit prompt that catches AI-generated text slipping into a generic tone. Checking instead of rewriting. A weaker model can do that too.

Not the tool, the combination

The debate about which coding tool is the best one misses the point. Claude Code writes the best code but checks itself badly. Codex is a strong auditor but not my go-to for building. Antigravity researches deeply, OpenCode handles the routine.

The real strength is not in any single tool. It is that they all reach into the same context and that I can reassign the roles as soon as the tools move on. Without my vault every switch would be a cold start. With it, it is a system.