[1/2] OpenClaw Creator: Why 80% of Apps Will Disappear (Category AI)
An interesting short interview with Peter Steinberger, creator of OpenClaw, who had recently joined OpenAI. He discusses local-first personal agents and how they could absorb much of the software we use. The interviewer is Raphael Schaad, a Visiting Partner at Y Combinator and founder/CEO of Cron, which Notion acquired. Here are the main ideas.
1️⃣ A local-first agent is more than another chatbot It runs on your machine or server, linking chats to your tools: files, shell, browser, and integrations—anything a user can do on a computer. That offers fundamentally different capabilities from a cloud assistant. Amusingly, the creator of Manus AI described moving in the opposite direction, from a local browser agent to the cloud to remove dependence on a local machine :)
2️⃣ The magic is tool use and finding the shortest route The most revealing example is a voice message that was transcribed even though nobody had explicitly built that feature. The agent:
- Identified the audio format from its header, even without a file extension.
- Converted it with ffmpeg.
- Chose not to install Whisper locally, instead using curl to send the audio to a speech-to-text API because that would get a result faster. Steinberger argues that competitive advantage is moving from the smartest model toward tools, orchestration, and the user’s data and memory.
3️⃣ “80% of apps will disappear” means apps that act as data containers The claim is not that UI will die. Apps whose main purpose is entering, storing, or managing data become redundant if an agent can:
- Gather context from files, logs, and photos.
- Read and write storage.
- Plan and remind.
- Deliver the result without a separate trip into an app. He separately suggests that products closely tied to specific hardware or sensors will survive.
4️⃣ Agent memory is both the main value and the main risk A personal agent’s memory—habits, context, action history, and local files—becomes a new private boundary. These memory files may be more sensitive than search history. Files such as SOUL.md, AGENTS.md, and TOOLS.md are also discussed as context injected into the workspace.
5️⃣ The CLI as a universal agent adapter One engineering idea runs against the current trend: agents often need ordinary human command-line tools rather than specialized integration protocols. With --help, examples, and predictable commands, an agent can often work out what to do. The interview presents this as a practical direction, including a discussion of CLI versus MCP.
6️⃣ Security from the start An agent that reads chats and the web while accessing shells, files, and browsers greatly expands the attack surface. The discussion emphasizes basic controls: threat modeling, isolation, allowlists, auditing, sandboxing, direct-message policies, and protection against prompt injection.
In the next part, we will discuss what engineers and technical leaders can take from this interview.
#AI #Engineering #Software #Management #Leadership #Startup #LLM #ML #Architecture