A Zoo of Agents
I have run a few coding agents side by side like Copilot, Codex, Claude Code, etc. Underneath they are the same kind of loop over mostly the same models, and feeding the same harness the same task hands you the same result. That part is settled. Let's move on.
So where does the disagreement come from? It comes from fit. Ask your friends they will give you different answers. Joe swears by Codex because it lands exactly right in their hands. Sarah reaches for the same capability inside Cursor and it grates. The tool did not change but the person did. A harness is a pile of "defaults" and habits and small frictions, and those either match how you think or they get in your way.
Which means the harness has to be personal. The files matter, and so does the name, the shape, the tools it reaches for, the way it talks back to you. It has to feel like yours. When it does, you get the WOW moment. No amount of raw intelligence underneath will save it when it does not fit.
This is why Anthropic dropping one agent called Claude into every Slack channel bothers me. It is a clever launch, and it is also one personality standing in every room. When Claude fumbles in a channel, the person on the other end remembers it, and carries that impression into a different channel doing a different job. The failure travels. The model is barely involved - what sticks is how that one harness felt in that one moment.
We learned this at CBK.AI. There you build a specific agent by choosing the components and how they fit together. We run twelve of them across the business now, and I have started building my own. They have different names and exist for different reasons. Some have a space to keep files, some hold shell tools and can drive a computer and there are those that are deliberately small. They each feel very different. I like some and not others, and that is the fun part - the ones I dislike never drag down the ones I love, and every so often I stumble onto a real gem. The ones that I don't love will possibly get depricated and replaced if the team also does not love them too.
None of them need the same components. And they should not. That difference is what makes the whole thing feel alive.
A Slack full of one Claude is a quiet, undifferentiated room. A workspace full of ten different agents is a zoo.
I surely know which one I want to walk into.