Peter Steinberger is the creator of Clawdbot (it has been renamed to Moltbot on January 2026) and founder of PSPDFKit.

Moltbot is a work-in-progress AI agent that shows what the future of Siri could be like. It is currently the hottest AI project in the tech industry, with more searches on Google than Claude Code or Codex in the begining of the year 2026.

Peter’s background is fascinating. He built and scaled PSPDFKit into a global developer tools business. Then, after a three-year break, he returned to building. This time, LLMs and AI agents sit at the center of his workflow.

Here are some highlight of the experiences of Peter Steinberger when building Moltbot :

1) Managing a dev team teaches you to let go of perfectionism : a skill important when working with AI agents. Running PSPDFKit with 70+ people forced Peter to accept that code wouldn’t always match his exact preferences. This makes him more efficient when working with agents today.

2) Close the loop : AI agents must be able to verify their own work. Peter designs systems so agents can compile, lint, execute, and validate output themselves.

3) Pull requests are dead, long live “promot requests.” Peter now views PRs as “prompt requests” and is more interested in seeing the prompts that generated code than the code itself. Interestingly, this is exactly what my brother, Balint Orosz said when he explained that they reject almost all external pull requests from Craft Agents, but take the core idea and use them as prompts later.

4) Code reviews are dead for this workflow—architecture discussions replace them. Even in Discord, he doesn’t talk code with his core team : they only talk about architecture and big decisions.

5) Runs 5-10 agents and stays in the “flow” state. Peter queues up multiple agents working on different features simultaneously.

6) Spend a lot of time planning out the work the agent will do, and prefers using Codex. Peter spent a surprisingly long time going back-and-forth with an agent to come up with a solid plan. He challenges the agent, tweaks it, pushed back. When he is satisfied with the plan, he kicks it off, and moves on to the next one. He likes using Codex because Codex goes off and does long-running tasks : Claude Code comes back for clarifications, which he finds distracting — given he fleshed out a plan already.

7) Under-prompt intentionally to discover unexpected solutions. Peter sometimes gives vague prompts to let the AI explore directions he hadn’t considered.

8) Local CI beats remote CI for agent-driven development. Peter runs tests locally through his agents rather than waiting for remote CI pipelines. He does this because he doesn’t want to wait an extra 10-ish minutes for a remote CI to run, when his agents can run tests locally.

9) Most code is boring data transformation—focus energy on system design instead. Peter argues that the majority of application code is just “massaging data in different forms” and doesn’t warrant obsessive attention.

10) Engineers who thrive with AI care about outcomes over implementation details. Peter observes engineers who love to solve algorithmic puzzles to struggling going “AI-native” like he has. People who love shipping products, on the other hand, excel.

Clawdbot is exploding across tech Twitter and Hacker News, with builders racing to plug autonomous agents into their lives before anyone really knows what they’re doing. Claire jumped in at the peak of the hype to find out what’s real, what’s risky, and what happens when you actually let one of these agents run loose on your computer.

Here are some highlight of the experiences of the user Claire when using Moltbot :

A) Security should be your top concern. Claire created a separate user account on her laptop for Clawdbot, gave it its own email address rather than access to hers, and created a limited vault in 1Password. When it requested broad permissions to her Google account (email, contacts, files), she pushed back and limited access to just calendar viewing. These precautions are essential, since Clawdbot has access to your file system.

B) Prompting matters more than ever with autonomous agents. Small differences in how you phrase requests can lead to dramatically different outcomes with Clawdbot. When Claire asked it to email podcast guests, she didn’t explicitly say “draft an email for my review,” resulting in immediate sending. With traditional AI tools, this wouldn’t matter as much, but autonomous agents require more precise instructions.

C) Clawdbot is biased toward impersonation rather than acting as an assistant. When asked to email podcast guests about rescheduling, it sent emails as Claire rather than identifying itself as her assistant—despite Claire explicitly giving it its own identity. This suggests that the default behavior is impersonation, which creates significant risks if not carefully managed.

D) The best use case might be asynchronous research tasks rather than real-time assistance. Claire’s most successful experience was asking Clawdbot to research Reddit for feedback about her product. The AI produced a well-organized report with key insights and reference links, exactly what she’d expect from a human research assistant.

E) The future of personal AI assistants will require both better interfaces and better security models. While Clawdbot demonstrates the potential of autonomous agents, it also highlights the challenges. The ideal solution would combine the accessibility of consumer products with proper security boundaries, clear identity management, and reliable performance—a combination that doesn’t yet exist in the market.

F) The “one-line install” is misleading. Expect to spend hours on setup if you’re not a developer. Claire spent two hours installing dependencies before she could even run the one-line installer. She had to upgrade Node, install Homebrew, install Xcode, and update NPM manually. This makes Clawdbot firmly in the “hacker/tinkerer” category rather than consumer-ready.

G) The latency problem makes Clawdbot feel slow compared with other AI tools. Unlike ChatGPT or Claude, which provide real-time feedback as they work, Clawdbot often goes silent while spinning up subagents to complete tasks. This creates a frustrating user experience where you’re not sure if it’s working or has crashed. Claire had to specifically ask for acknowledgment messages, which it still didn’t consistently provide.

H) Clawdbot struggles with basic time concepts despite its advanced capabilities. In a particularly frustrating episode, Clawdbot placed all family calendar events on the wrong day and couldn’t set recurring events. When Claire pointed this out, it claimed it was “mentally calculating” which day of the week each date falls on—a bizarre explanation for an AI.

I) Voice plus text messaging creates a surprisingly natural interface for an AI assistant. One of Clawdbot’s strengths is its ability to receive voice messages via Telegram and respond with either text or voice. Claire found this particularly useful while shopping at Target, allowing her to multitask while managing the AI. The ability to instantly add this capability without complex setup demonstrates the power of self-learning agents.

J) Clawdbot is not just for Mac Minis—it can run on any machine or in the cloud. Despite what many believe, you don’t need special hardware to run Clawdbot. Claire ran it on an old MacBook Air, and you can even set it up for $5 on Amazon Web Services. The key is understanding that while it runs locally, it doesn’t require fancy hardware unless you’re running massive local models.