Why we watch him. Peter Steinberger is a rare first-party builder's voice on autonomous agents rather than a commentator's. He shipped a runaway open-source agent, watched it pass 180,000 GitHub stars, then made a real governance call about it (an independent foundation, not an acquisition) while joining OpenAI himself. That combination, hands-on agentic engineering plus a live view of where agents replace apps, sits close to Dominic's core coding-vs-knowledge-work theme, and complements Boris Cherny (coding agents at Anthropic) and Aaron Levie (enterprise AI) from the personal-agent side of the frontier.
Current headline view: software is moving from apps you open to a single agent you ask, and that agent should be able to read, adjust and rewrite its own tooling as it works. He expects agents to make most consumer apps irrelevant within a few years; survival depends on offering a clean API an agent can call. He also draws a hard line between disciplined "agentic engineering" and casual "vibe coding".
Latest 1-page summary (as at 10 Jul 2026) 🟠 new thinking
The six core theses we track him against:
Self-modifying agents. OpenClaw can introspect its own source, read its own docs and tool list, and rewrite parts of itself when something is not working, treating self-modification as a debugging strategy rather than a one-off feature.
Agent self-awareness as a design choice. He deliberately made the agent aware of what it is, its own source, where its docs live, which model it runs, so it can reason about its own behaviour rather than operate as a black box.
Agentic engineering beats vibe coding. A real discipline (context-window management, codebase structure, feedback loops) is distinct from, and more durable than, casually prompting and hoping. He describes a U-curve: simple prompts, then overcomplicated multi-agent orchestration, then back to simple prompts with real understanding underneath.
Empathy over micromanagement. The best people working with agents are not necessarily the best coders, they are the ones who can empathise with a system that starts fresh with no memory every session.
Agents will kill most apps. He expects roughly 80% of consumer apps to become irrelevant as people default to asking their agent rather than opening a dedicated app; survival depends on a clean agent-callable API.
Open source as the right home for a runaway project. Despite acquisition interest from major labs, he moved OpenClaw to an independent foundation to keep it open, while separately joining OpenAI himself, treating his own career and the project's governance as separable decisions.
Method note. This v1 baseline is distilled from press coverage of the Lex Fridman #491 interview and his own site (10 Jul 2026), not yet a full transcript or X history read. The weekly scan refines it from here and flags any deviation.
Analysis backlog (newest first)
Deep source docs are held locally in the workspace at expert-watch/, not on this page.
10 Jul 2026 · v1 profile built (from cact #66's own worked example, the Peter Steinberger podcast Dominic flagged as needing an "explain like a 15 year old" pass). Identity verified (steipete, OpenClaw/Clawdbot creator, ex-PSPDFKit). Six core theses distilled from press coverage of Lex Fridman #491 and his site.