OpenClaw founder Peter Steinberger on riding the personal agent hype cycle, agent wars, and why he's now excited about Linux

Peter Steinberger, OpenClaw

Peter Steinberger built OpenClaw out of frustration. He couldn't talk to his computer the way he wanted to, found the existing workflows inadequate, and started exploring alternatives. The result was one of the earlier demonstrations that personal AI agents could do more than summarize email — they could install software, make purchases, and act proactively on a user's behalf.

The hype cycle, honestly

Steinberger is candid that OpenClaw rode a hype cycle that outran the product. Expectations inflated well beyond what a solo builder could deliver, and when users discovered the limitations, the backlash was predictable. He was also unapologetic about the vibe-coded origins of the project, arguing that trusting the models and shipping fast was the right call even when critics called it slop. The models kept improving, and the paradigm he was pointing at kept validating itself.

What he thinks actually drove OpenClaw's impact wasn't a specific feature set but a demonstration effect: he showed people a different way to use AI, packaged in something that felt less like a wall of text and more like a live, proactive collaborator.

“I just showed people a different way how to use AI. So many were stuck in this world, oh, it can summarize my email. No. It can install stuff on your computer or buy things for you or do whatever you want that you can do in your computer if you just only give it the right tools... We had a massive hype cycle with inflated expectations of what Open Cloud would do for you. And then of course people tried it, they're like, no, it doesn't quite work like this.”

Agent wars

The competitive dynamic Steinberger sees forming is companies blocking third-party agents to push their own. His instinct is that this is a losing strategy. If a service blocks his agent, he buys somewhere else. The better move, in his view, is making your service easy for any agent to use rather than trying to own the agent relationship.

He frames the likely resolution through a lottery-style model: instead of whoever-clicks-fastest winning a limited-supply product drop, systems should allocate access in ways that are agent-neutral. The old first-click-wins paradigm breaks down when agents can click in 0.1 seconds.

Personal versus army

On the open question of whether the future is one personal agent or a coordinated fleet, Steinberger leans toward the personal agent model but acknowledges the multi-agent hierarchy is equally plausible. He's been experimenting with agents that spawn sub-agents on demand, and describes wanting to build something where a single instruction — say, rewriting Postgres in Rust — triggers automatic planning, team structure, and timeline management without further input.

One experiment that produced real results: he moved his entire team to a shared workspace where everyone's agent prompts are visible. The initial reaction was embarrassment — prompting felt private. Within a day, it became a learning accelerator. Teams that had built completely different mental models for working with agents suddenly had visibility into each other's approaches.

Why Linux is interesting now

The frontier Steinberger is most excited about is the operating system itself. As models get better, he's found himself annoyed that he can't modify software through a prompt. Linux solves that. His agent can read the kernel, compile it, and fix driver issues — he describes gluing a webcam to a display so his agent could see what was happening, running it overnight, and waking up to a fixed display driver. The audio fix took longer because the agent spent the night testing volume levels before realizing the output was simply muted.

The underlying thesis is that nothing is impossible anymore, just hard. Running a custom OS is still a significant undertaking, but it's no longer an absurd one.

What he's building now

Steinberger describes splitting time between OpenClaw work and contributing on the foundation side at OpenAI, with cross-pollination in both directions — taking ideas from OpenAI back to his own projects, and feeding product experiments back in. The through-line is the same question he started with: as models improve, how much further up the stack can software move?

His working answer is that whatever workflow exists today is already outdated. Models are advancing faster than anyone can build interfaces around them, and that gap is where the opportunity sits.