Skip to content
#AI

[1/2] OpenClaw Creator: Why 80% Of Apps Will Disappear (AI column)

#AI #Engineering #Software #Management #Leadership #Startup #LLM #ML #Architecture

Interesting and short. interview Peter Steinberger, creator of OpenClaw (recently moved to OpenAI) Prolocal-first personal agents and how they can "eat" most of the usual software. Interviewed by Raphael Schaad, Visiting Partner at Y Combinator; Founder/CEO Cron (The project bought Notion). Below are the main insights

1Local-first agent Another chatbot The agent lives on your machine/your server, so it can be a gateway between chat rooms and your tools: files, shell, browser, integrations - "anything a user can do on a computer." This is a different class of capabilities compared to the cloud assistant. Funny how the creator of Manus AI told about how they left the local agent in the browser to the cloud to untie the local machine:)

2Magic is a tool-use + shortcut The most revealing case in the interview was that the voice message was suddenly transcribed, even though the functionality was “clearly not encoded in advance.” The agent himself made the next shig. Identified the audio format by title (without expansion)

  • Over the ffmpeg. I did not put locally Whisper, but chose the quick way - to send audio to the speech-to-text API via curl, because it is faster to get the result.

As a result, the competitive advantage shifts from the “smartest model” to tools, orchestration and user data/memory.

3️⃣ "80Percentage of applications will disappear: Data container applications The point of the thesis is not that the UI will die. Applications whose primary function is data input/storage/management become redundant if the agent:

  • He's collecting context. (file/log/photo)
  • Writes/reads from the vaults
  • Planning/reminding And does the result without a separate “go to the application”.

Separately, the idea that products strongly tied to specific iron / sensors will survive.

4The main value and risk - the "memory" of the agent If an agent becomes personal, then their “memory” – habits, context, history, local files – is the new private perimeter. The report noted that such memory files may be “more private” than conventional search history. Files like SOUL.md/AGENTS.md/TOOLS.md are also mentioned as part of the “injectable” context in workspace.

5CLI as a universal adapter for an agent One of the engineering ideas, which contradicts the current vector: instead of special integration protocols, agents often suffice with ordinary human tools - CLI. If you have --help, examples, and predictable commands, the agent will often "find out." In the report, this is presented as a practical vector. (Including discussion of CLI vs MCP).

6Security - not "after", but immediately An agent that reads chats/web and has access to a shell/files/browser dramatically expands the attack surface. The report focuses on basic controls: threat model, isolation, allowlist, audit, sandboxing, policies for DMs, protection against prompt injection.

In continuation We will discuss what engineers and technical managers can learn from this interview.

#AI #Engineering #Software #Management #Leadership #Startup #LLM #ML #Architecture