TLDR:
Generic agent utility is limited
Deep personalization makes agents more useful and more sticky
IKEA effect: Power users value what they configure themselves
Safety vs utility tradeoff always has a clear favorite
OpenClaw was an early signal, not a one-off
Dispatch shows the category is already getting packaged for a much broader audience
Big tech already saw the opportunity
Ever since it burst the AI bubble, the conversation around OpenClaw, the autonomous AI agent, split into two camps. One camp saw a security disaster that should not exist. The other saw the magical personal chief of staff that would automate their whole life and maybe make them money on the side. Both camps miss the point entirely. OpenClaw proved market fit, Claude Cowork Dispatch is now the mainstream confirmation.
On the one hand, it's fair to say that security issues were obvious from day one. If you wire an agent into code execution, messaging, and personal data, people are going to freak out for valid reasons. OpenClaw was never a polished enterprise product. It was a playground project that got dragged into the center of the timeline, built by Peter Steinberger to see how far a personal agent could go if you fully leaned into code, CLI tools, and user control. The guy was just having fun messing around with agents.
Generic agents are useful, but limited. OpenClaw lets power users build an agent around their own tools, own data, own habits, and own taste. The IKEA effect kicks in hard here. Power users trust what they configured themselves, and with agents the configuration is half the product. And it's addictive. You're literally building each piece of it, almost like building an RPG character.
Judging by the repo growth, the number of forks, and the pace of adjacent projects and the community around it, this looks like a clean demand signal. People do not just want an AI assistant. They want their AI assistant.
OpenClaw's record
That spread is another tell. OpenClaw feels like this year's DeepSeek moment: one open project breaks containment, China picks it up hard, and local variants start multiplying. Kimi Claw rode the wave to #2 on the February product growth rankings, Tencent pushed QClaw, and suddenly this looked less like one viral repo and more like a whole category.
Once the format caught on, claw variants spread fast enough to show up in the rankings.
OpenAI saw that quickly. Peter announced he is joining OpenAI to work on bringing agents to everyone. Whatever happens there it doesn't matter, because the idea already escaped the repo.
Now, NVIDIA moved in too. NemoClaw takes the OpenClaw idea and wraps it in enterprise language, privacy controls, and security rails.
NVIDIA's wants to keep selling their shovels
Anthropic joins the game
Anthropic started to move in that direction with products like Cowork, which feels like an attempt to bring agentic tooling to people who are not developers. And just as I was writing this piece they released.
Claude Cowork as a research preview that I'm excited about: Dispatch! One persistent conversation with Claude that runs on your computer. Message it from your phone. Come back to finished work.
It feels pretty magical to give Claude a mission on my computer and getting occasional updates, like creating reports from internal dashboards or finding me a better seat on my next flight.
Funny how nothing fundamentally changed in the safety model. I wonder how this rollout will work out.
Why I think this is here to stay
Agent utility does not stop at coding. Coding is just where this clicked first because the users are technical and the feedback loop is tight. Once you can create skills for different use cases, you start seeing how this expands into everything else too. Research, admin, personal ops, knowledge work, whatever. Maybe this is the dawn of a new computing platform.
The agent platform pitch
So this is still just the beginning. The utility only really starts skyrocketing once the agent is plugged into your own data, your own knowledge, and your own workflow. That is a huge part of the OpenClaw appeal. The other big part is remote control. Being out in the world, talking to your agent, and having it do actual work for you back on your machine is insanely useful.
That is why I think this might become a real computing shape. A customizable agent wired into your apps and personal context. And if you add something like a skill (or the next iteration of that) marketplace, where you can share and install skills, the whole package gets too good to avoid.
My hunch is that this spreads more like an ecosystem than a single app. We probably end up with one or two major OpenClaw-style operating systems, if that term even makes sense, and then a bunch of different flavors around them, something like Linux distros for agents.
Elsewhere in the Latent Space 🌌
Claude 4.6 models now include the full 1M context window - Although it seems to score better than others, I'm still very skeptical about this. Anything past 100k feels like dangerous territory.
OpenAI's Symphony showcases the idea of a ticket driven approach for software development. Lovely. The funny part is how they pulled that off. It is a cron job dispatching tickets from Linear to workers, each of which is a Ralph loop using a Linear comment as draft pad for persisted state. Says a lot about the state of agentic coding tooling.
Kevin Gu breaks down why files might not be a good enough solution for solving agentic memory. It works fine with a low amount of files, but at a company level, it’s utopian. You won’t avoid the drifting challenges and the performance issues.