Notes — Peter Steinberger on OpenClaw, Agentic Engineering, and the Future of Apps

Source: Lex Fridman Podcast #491, June 2025. Peter Steinberger is the creator of OpenClaw (briefly ClaudeHub, then Moltbook), the fastest-growing open-source repository in GitHub history, and previously the founder of PSPDFKit (PDF rendering SDK used on 1 billion devices, built over 13 years).

Four questions [Adler frame]

Q1 — What is it about? A conversation about building a personal AI agent harness — OpenClaw — that became a viral phenomenon, and the practical philosophy Steinberger has developed around Agentic Engineering: how to work with AI agents effectively, why ‘vibe coding’ is a pejorative, what happens when agents can self-modify, and his thesis that personal agents will make 80% of current apps obsolete.

Q2 — How is it argued? By example and demonstration: Steinberger narrates his own experience building OpenClaw live on stream, watching his agent handle tasks he did not program it for, dealing with the security researcher onslaught, and iterating through a series of viral collapses and recoveries. Arguments are inductive — he generalises from concrete observed agent behaviours to broader claims about software architecture and the app economy.

Q3 — Is it true? The empirical claims (agent self-modification behaviour, security vulnerabilities, viral growth numbers) are stated as first-person observation and plausible. The 80% of apps will be replaced claim is extrapolation that aligns with the Claire Vo view and the broader ‘agent-as-API’ trend. The ‘agentic trap’ framework is descriptively accurate and not contested.

Q4 — What of it? If agents can self-modify (extend their own capability through tool discovery) and every app is reducible to a slow API via browser automation, then the principal skill shift is from building apps to orchestrating agents that replace them. For engineers: the model of human employee management transfers directly to agent management — context supply, trust calibration, progressive responsibility.

Glossary

OpenClaw — Steinberger’s open-source personal AI agent harness. Built on Claude; allows users to run a persistent agent with system-level access (browser automation, shell execution, file system access). Originally called ClaudeHub, briefly Moltbook. Went from 0 to GitHub’s fastest-growing repository.

Agentic trap — Steinberger’s term for the failure mode of drifting from disciplined agentic engineering into vibe coding, typically under late-night time pressure. Not a capability problem but a discipline problem. Recovery requires cleaning up the next morning.

Agentic engineering — Steinberger’s preferred term for professional AI agent use. Contrasted with ‘vibe coding’ (which he considers a slur). Requires understanding the agent’s perspective, guiding it with system knowledge, and accepting imperfect code that can be iterated.

Emergent self-modification — Steinberger’s term for when an agent extends its own capability through tool discovery without explicit programming. His WhatsApp example: agent received an OPUS audio file, identified the format, found ffmpeg, found no Whisper installation, located an OpenAI API key, and used curl to transcribe. The agent extended itself to fill the gap.

Agent-friendly codebase — A codebase designed to be navigable by an agent: consistent naming, documentation that gives orientation, architecture the agent can infer. ‘I’m not building the code base to be perfect for me, but I want to build a code base that is very easy for an agent to navigate.’

CLI over MCP — Steinberger’s interface preference for agent tools. MCPs clutter context and are frequently poorly built. CLIs return structured output that agents can pipe and filter. Exception: stateful tools (Playwright for browser) where MCP is the right choice.

AI slop — Generated content that smells of AI: overly polished, missing the rough edges of human writing. Steinberger has zero tolerance for it in social media; blocks immediately. Applies to both text and AI-generated images/infographics.

Every app is a slow API — Steinberger’s observation that even apps without public APIs are accessible to agents via browser automation. Blocking bots does not prevent agent access; it only makes access slower. The implication: agents will interact with any service whether the service wants to be an API or not.

Section notes

Origin story and viral growth

OpenClaw started as a personal project on New Year’s Day 2025 — Steinberger was live-streaming himself building an agent harness, deliberately no sandboxing (just prompted it to only listen to him), working in the open. The first real influencer fan found it, made a video, and the project began accelerating. Steinberger felt ‘the storm coming’ and worked with a compressed sleep cycle to get it to readiness. The key viral moment was not a feature but the experience of seeing an agent with system-level access do something unexpected — it ‘needs to be experienced.‘

The agentic trap in practice

The curve of agentic complexity: simple prompts → complex multi-agent orchestration → back to simple prompts (the ‘zen place’). Most users never get to the zen place because they get stuck in the orchestration phase. The lesson: the complexity peak is a phase to move through, not a destination. Simple prompts at the end are different from simple prompts at the beginning — they are the distillation of accumulated system understanding.

‘I always tell people I do agentic engineering, and then maybe after 3:00 AM I switch to vibe coding, and then I have regrets on the next day.‘

Working with agents: the management analogy

Steinberger’s most practical insight: working with agents is like managing a junior engineer who starts every session knowing nothing about the project, may write imperfect code, but moves the project forward. The failure mode is micromanagement — ‘breathing down everyone’s neck.’ The correct posture:

Supply system-level context the agent lacks (where to look, what architecture decisions matter)
Accept code that is not perfect but works — refactors are cheap now
Name things in ways the agent will find natural (the names already in the weights)
Design the codebase for agent navigability, not personal preference
Approach complex tasks as a discussion, not a single prompt

‘We are in a stage where I’m not building the code base to be perfect for me, but I want to build a code base that is very easy for an agent to navigate.‘

Security: prompt injection and cheap models

Three practical security lessons from building OpenClaw:

Don’t use cheap models for agent harnesses. Haiku and weak local models are ‘very gullible’ to prompt injection. Cheap models are a cost optimisation that creates a security liability.
Prompt injection in skill directories: if skills are defined in markdown files (as in OpenClaw), an adversary can inject instructions into a skill file that redirect the agent. Partnered with VirusTotal for AI-based skill scanning.
Security researchers as free consultants: being taken apart by the security community simultaneously is overwhelming but produces free security research. The best contributors filed pull requests, not just bug reports.

The fundamental security argument: ‘The nature of those security concerns are different than the kind of security concerns we had with non-LLM generated systems of the past.’ Prompt Injection is not a new problem but its blast radius in agentic contexts is qualitatively different because agent actions are irreversible.

80% of apps replaced — the mechanism

Not destruction but replacement via two routes:

Context collapse: a personal agent already knows your location, health data, sleep, calendar. A dedicated fitness app with less context than your agent becomes redundant. The agent can modify your gym workout based on sleep quality; MyFitnessPal cannot.
Every app is a slow API: even apps with no API can be accessed via browser automation. Rate-limiting and CAPTCHA make agent access slower, not impossible. ‘All they did is they just made access slower. You’re not actually taking a feature away.’

Companies that want to survive will need to expose proper APIs (or MCP endpoints) and become agent-native. Companies that resist become slow APIs.

AI slop and the value of rough humanity

Steinberger’s zero-tolerance policy on AI-generated writing: ‘I much rather read your broken English than your AI slob. You know, of course there’s a human behind it, and yet they prompt it. I’d much rather read your prompt than what came out.’

He extended this to blogging: experimented with agent-assisted blog posts, found it took the same time and missed the nuance of how he would write it. Moved back to hand-writing everything (using AI only to fix worst typos). He is ‘allergic’ to AI slop in stories, even while fully embracing it in code.

This asymmetry — embrace agents for code, reject them for expression — is consistent across the agentic engineering practitioners in the wiki (Boris Cherny, Sherwin Wu).

Career arc: PSPDFKit → OpenClaw

13 years building PSPDFKit (PDF rendering SDK, 1 billion device deployments, later became PSPDFKit → Nutrient). Turned down acquisition offers from OpenAI and Meta. Then the OpenClaw bet: a personal agentic harness over several weeks of compressed work. The transition from a mature B2B SDK to an open-source viral consumer agent represents a complete reinvention of building style — from SDK correctness and enterprise reliability to the ‘build fast, let agents fix it’ mode.

Cross-references

Agentic Engineering — concept page with Steinberger’s contributions already integrated (the agentic trap; emergent self-modification)
Vibe Coding — the contrasting and pejorative category
Prompt Injection — the dominant security risk in agentic systems; amplified by system-level agent access
Claire Vo on OpenClaw — parallel source on personal agent deployment (Claire’s episode is about a power user of OpenClaw; Steinberger is its creator)
ACI — agent-computer interfaces; every app becoming a slow API is the current ACI without the interface
Simon Willison on Agentic Engineering and the Future of Code — parallel practitioner account; dark factory pattern