Is Agent Hopping a Thing?
If you’ve ever used Linux, you know the feeling. You install Mint one weekend because it’s friendly and just works. Then Ubuntu catches your eye — bigger community, more packages. Pop!_OS promises a polished dev experience out of the box. Kubuntu lets you tweak every pixel. You tell yourself this is the one. Then a month later you’re reading wiki pages at 2 a.m., wondering if you should give the next distro a shot. That’s distro hopping.
And lately I caught myself doing the exact same thing. Just not with operating systems. With AI agents.
Turns out I’ve been pulling the same move — bouncing between platforms that each nail one thing but fumble another. ChatGPT got me hooked first, but it’s a walled garden. Continue was open and customizable, yet the context search never clicked. Windsurf had all the agentic coding goodies but locked me into their models. Mistral gave me API access bundled with a single subscription — nice for experiments. OpenCode finally felt like home. And then I stumbled onto LobeHub and discovered remote device execution, which I hadn’t seen anywhere else, and suddenly the whole game changed again.
The dream is a workflow that just works everywhere — the kind where you can walk with your son and steer an AI from your phone, because there just aren’t enough hours in the day. The only thing I envy AI? It doesn’t need to sleep. So here I am, chasing that setup. And if any of this sounds familiar, maybe you’ve been agent hopping too.
And if you’re wondering whether this is just me — it’s not. In a recent survey by The Pragmatic Engineer, 70% of engineers reported using between two and four AI coding tools. Only 15% stick with a single one, and another 15% juggle five or more. We clearly haven’t settled on a single tool yet, and maybe that’s the whole point.
Why Not Claude Code, Codex, Copilot?
It would be too easy, wouldn’t it? These are market leaders for a reason. But that also means they’re built to fit the majority’s expectations — just not mine.
I want to actually understand what I’m working with. Open-source tools let me see the stack, learn the internals, and adapt when things change. Sure, configuration takes more effort upfront, but I end up knowing my workflow inside out. Once you know your setup that well, the knowledge sticks — no matter what tool or job comes next. Nobody can guarantee your next gig will use the same provider.
Then there’s the lock-in thing. When you build your daily workflow around one vendor’s models and one vendor’s platform, switching gets painful. I like having the choice to swap things out without starting from scratch.
And yeah, I do have a vague, probably irrational discomfort with handing too much of my process to one big corporation. I just like to keep my options open.
History of Hopping
A quick disclaimer: this is my personal journey — not a comprehensive review. Every tool I mention here is genuinely amazing and has its place. This is just one guy’s path through the landscape.
ChatGPT was my first. I first learned about it in a Data Science class — and it felt like real intelligence. Free subscription, plain chat mode — and honestly, it helped a lot. At the time I didn’t know a thing about agents, system prompts, web search, and all the other great stuff that would improve raw LLM intelligence in the years to come. It was enough to get me hooked, but it stayed more of a conversation partner than a coding companion.
That’s what led me to Continue. Open-source, highly customizable — exactly what I wanted on paper. But I struggled with getting relevant context search to work right. It probably wasn’t a bad option, but the bumpy start didn’t inspire confidence, and I moved on quicker than I probably should have.
Windsurf came next — and it had all the goodies of agentic coding. But the cracks showed fast: a limited list of models and providers, still vendor lock-in, still a roomy prison. It wasn’t snappy in IntelliJ, and there was no CLI or web option. The walls felt too close.
Then I tried Mistral Vibe. Good models, and a real plus for EU digital sovereignty. The subscription comes with API access, which is genuinely nice for custom agentic experiments — I could plug it into different coding agents. The downside? Sometimes I needed a more powerful model than what was available, and the limits felt lower when using OpenCode with the API instead of the native Vibe CLI. Good concept, but the juice wasn’t always worth the squeeze.
And then things started clicking.
OpenCode was the first one that actually grew on me. Customizable, powerful — agents, skills, any external provider. Lightweight and snappy. Terminal, desktop, IDE integrations — pick your flavor. And Docker sandboxes support it out of the box. It’s popular for a reason. I’m still exploring everything it can do.
LobeHub was an unexpected discovery. I have mixed feelings — I failed to get an OpenAI-compatible setup working, and failed to get Mistral working too. But it covers desktop, web, and mobile, and the remote device feature is a genuine killer feature. Seriously, I haven’t seen anything like it anywhere. Here’s how it works: you keep the CLI running on your machine and everything goes through WebSocket. You ask the agent from your phone, it executes on your laptop. Of course you need the sandbox for this. But that’s the kind of feature you didn’t know you needed until you see it work. Probably possible with OpenCode too if you set up a server — but that’s environment setup, and you’d need something more powerful as your server if you want parallel agents spinning up containers and running tests. That’s going to cost you, all while your personal machine with plenty of RAM and processing power sits idle or sleeps. With LobeHub, that’s not the case — it can run the agent on your machine and report back to the Hub. As for the planning and orchestration side, I’ll be honest: I still haven’t fully grasped the whole thing yet.
So where am I now? Still exploring. Still hopping. OpenCode feels like home base for coding, LobeHub keeps surprising me, and I haven’t sworn off any of them permanently. The quest continues.
Reflect on Expectations
When I first started this journey, I thought I was looking for the best coding agent. And sure — coding excellence matters. The context and codebase understanding, working with git, following all the guidelines a real developer would follow. OpenCode impressed me there: customizable, powerful, any model provider I want behind it, granular control over what tools run and what they can touch. Permission and security I can actually configure. That matters. A lot.
But somewhere along the way the checklist grew legs. I caught myself asking: can I task the agent from my phone when I’m away to do development or research on my laptop back home? Can I schedule something to run overnight? LobeHub’s remote device execution was an eye-opener there — the kind of thing you don’t know you need until it works. And it got me thinking bigger.
Then there’s the architecture I’m slowly picturing. Building complex agent architecture on diverse models — plan on the expensive ones, spawn subagents on cheaper ones. Let Subagents pollute their context windows with logs reading and test-fixing iterations, while the Supervisor stays clean, overseeing the progress. Run things on a schedule, trigger on PR comments. Build a workflow, a framework, a team. Both OpenCode and LobeHub seem capable of this, but I haven’t fully grasped the orchestration layer yet. That’s what I want to figure out next.
And underneath all of it: I want to keep the number of subscriptions low, clear what I pay for, predictable. Not a separate bill for chat, for the coding agent, for extra models, and god knows what else. I also want the safety net of open source. Not because I plan to self-host everything tomorrow — I don’t need to self-host everything. I’m more than happy to pay for nice and efficient cloud. But I need to know that the tooling I use, if the day comes, can be set up on premises, to comply with whatever regional restrictions or data safety policy might come up. That matters. It’s not paranoia — it’s just knowing the exit door exists.
The distro-hopping parallel holds up. Nobody hops distros to find the one that wins benchmarks. You hop because you’re figuring out what fits. What feels right under your fingers, what matches how you think and work. After all this hopping, I can finally say: I’m not looking for the best AI agent. I’m looking for the one that fits me — across devices, across models, with enough guardrails to feel safe and enough freedom to feel like mine.
I don’t know yet whether OpenCode and LobeHub will play nicely together, or if one will win out, or if something like Hermes will have it all in one. We’ll see. But the target is getting clearer. And for the first time, I actually know what I’m aiming at.