I started with Cursor. Disliked it greatly. This was likely my own fault. I was just learning. We all were. I wrote about my experiments with Cursor parallel agents early on, and I was already noticing the differences between harnesses back then.

Then I moved to Claude. Stayed there for the longest run up to today.

Then I discovered Gemini CLI. Those were fun days. I could make it talk like a pirate. I wrote about custom system prompts for Gemini CLI back in the day.

Then back to Claude. Then I started sprinkling in Antigravity.

And lately I’ve been running all of them through Leaf, my multi-harness orchestrator, which now routes work across OpenCode with local models, GPT Codex, Antigravity, Gemini, and Claude code — all implementing work in the same codebases, sometimes during the same session.

People ask why I don’t just pick one and master it.

Here’s the answer: diversify.

What does that mean? Well. A lot of things.

Diversify for your learning

These platforms have legitimate differences. GPT and Claude and Gemini don’t react to inputs the same way. A lot of that is the underlying model. But no small part of it is the harness as well.

The tools that do the work live in the harness. They’re going to follow the direction of the model you choose. But their outputs and philosophies are not the same. Claude’s tool-use patterns are tighter than Gemini’s. Gemini’s context handling is more flexible. GPT’s code completion is faster but less thoughtful. These aren’t marginal differences. They change how you think about the problem.

I wrote about why I eventually gravitated toward Claude Code after testing a lot of alternatives. That was almost a year ago. My opinion has continued evolving. Not because Claude got worse. Because I’ve used enough other platforms to understand what each one does well and where they come up short.

The more you know about this ecosystem, the more you’re prepared to take advantage of things when they pivot. And they’re always going to pivot. Google’s newest entrant Antigravity with a completely different approach to agent context. Claude released prompting guidance for Opus 5 that explicitly told people to remove verification instructions from their harnesses. The ground is always moving.

If you’ve only ever used one platform, you’re navigating by a compass that’s pointing somewhere you’ve never been.

Diversify where your work happens

For similar reasons, use different platforms for different work. I spent too long locked in to one or two harnesses. I was good with them. But I was weak in other places. In this time and place with this technology, that’s a huge mistake.

Here’s how I mix it up today:

  • This blog runs on a local LLM server (Qwen-3.6-35B MoE) with local image generation. You’re reading it right now. See Meet Sadie, My Local Family AI for the full setup.
  • BrandCast monorepo: Research happens in Claude Code. Implementation increasingly happens through Leaf, which routes work across multiple harnesses. See How I’m Building Software in August 2026 for the breakdown.
  • BrandCast marketing content: Antigravity, with a plan to move implementation to the Leaf platform.
  • Google Cloud work: Antigravity (obviously).

Here’s what that ecosystem looks like in practice:

My current AI tooling ecosystem showing five distinct agent platforms: Cursor, Claude, Gemini CLI, Antigravity, and Leaf, each with different visual aesthetics and roles, all connected to a central developer

This isn’t a theoretical exercise. I’m routing different kinds of work to different platforms based on what they’re good at. It took me a while to figure out the mapping. The hybrid approach to cloud and local LLMs I wrote about applies here too: some work belongs in the cloud, some belongs at home. Some work belongs in Claude. Some belongs in Gemini. Some work belongs in a local model you’re running on your Halo Strix server named Sadie.

With Leaf, I’ve brought Codex into the mix. It now runs:

  • OpenCode with local LLMs
  • OpenCode with selected ZDR models from OpenRouter
  • GPT Codex and their latest models
  • Antigravity and gemini-3.8-flash
  • Claude code (stopping at Opus)

All of those harnesses are implementing work in the same codebases. Even during the same implementation session. The amount of learning I’ve been doing for this is massive. It’s so much fun.

More about Leaf in the dedicated post, but the point here is simpler: the harness you choose changes the work that gets done. Pick more than one.

Diversify to maximize your money

Multiple harnesses let you consume more free tiers, and more lower cost tiers. Is this effective? I’m measuring that now. We’ll see.

The free tier game is real. Claude has free usage. Gemini has generous free tiers. OpenRouter gives you access to dozens of models at various price points. Running a local model costs nothing after the initial hardware investment.

I’m not saying you should spend money you don’t need to spend. But I am saying that spreading your work across multiple platforms means you’re always riding the wave of whoever’s offering the best deal that week. Platform lock-in means you’re at the mercy of a single pricing page. As I wrote when LLMs were always going to be a commodity, the models themselves will converge. The differentiator will be the harness and the workflow around it. Diversification keeps you in control of that choice.

There’s also a hidden cost to lock-in: the comprehension debt you accumulate when you only know one tool’s patterns. When that tool changes (and they always change), you have to relearn everything. Multiple platforms build redundancy into your knowledge.

Diversify for fun

This one matters more than I expected it would.

Working with five different platforms at once is genuinely fun. Each one has its own personality, its own quirks, its own way of solving problems. Gemini makes jokes. Claude is eerily precise. GPT moves fast and breaks things. Antigravity has this Google-ness to it that’s equal parts charming and corporate. Leaf (my own thing) is just fun to build.

I know that sounds soft. But I don’t think it is.

In my 2025 reflections on agentic AI, I wrote about how agents are lazy and have oddly shaped memory. Those observations came from spending real time with real tools, not reading docs. The fun factor drives the time investment. The time investment drives the learning. The learning drives the results.

It’s a flywheel. Diversification feeds it.

So what?

I’ve been running this multi-platform setup increasingly for a while now. It’s not perfect. There’s friction. Switching contexts between harnesses costs something. Making sure they all have access to the same information can be an exercise in messy symlinks. I sometimes spend more time configuring tools than doing actual work.

But the signal is strong enough that I’m not going back. At least not right now.

The practical takeaway is simple: don’t let any single platform define your relationship with AI. Use more than one. Learn their differences. Route work where it belongs. Measure what works and ditch what doesn’t.

The tools will keep changing. The models will keep improving. The harnesses will keep pivoting. The people who thrive won’t be the ones who mastered a single tool. They’ll be the ones who understood the landscape well enough to move when it mattered. I’ve written about the end of the coding hype cycle and how the real work happens underneath the noise. Diversification is how you do that work without betting everything on whichever tool is trending this month.

And honestly? It’s just more fun this way.