OpenClaw: The Open-Source AI Agent Sparking a Post-App Revolution
In late 2025, a small open-source project detonated across the internet like a well-placed charge under the foundations of the app economy.
Its name is OpenClaw.
What began as a self-hosted AI experiment rapidly became one of the fastest-growing open-source projects in history, amassing more than 160,000 GitHub stars in a matter of weeks. But OpenClaw is not merely another chatbot. It is something more radical — a local, autonomous AI agent that lives on your computer and turns everyday messaging apps into a command center for your digital life.
You don’t open OpenClaw.
You text it.
From WhatsApp, Telegram, Discord, Signal, Slack, iMessage, or even Microsoft Teams, you message your agent as if it were a friend — and it acts. It clears inboxes. Books flights. Manages calendars. Writes code. Browses the web. Controls smart lights. Summarizes conversations. Schedules meetings. Pays bills.
Unlike cloud AI assistants that respond politely but passively, OpenClaw has what its community calls “eyes and hands.”
It can:
Read and write files
Execute terminal commands
Control browsers
Take screenshots
Access your camera or location (with permission)
Run background tasks proactively
And by default, your data never leaves your machine.
In an era dominated by cloud monopolies, OpenClaw feels less like an app and more like reclaiming your digital sovereignty.
Origins: From Clawdbot to OpenClaw
The creator of OpenClaw is Peter Steinberger, an Austrian developer best known for founding PSPDFKit, a highly successful PDF SDK company. He describes himself, half-jokingly, as a “vibe coder” — someone who builds by intuition, momentum, and relentless iteration.
The project began modestly in November 2025 under the name Clawdbot, derived from a personal assistant called Clawd (a nod to Anthropic’s Claude). After a trademark complaint, it was renamed Moltbot — a playful lobster reference (Claude → claw → molt).
Three days later, Steinberger changed it again.
“Moltbot sounds awkward,” he said.
The final name stuck: OpenClaw.
In January 2026, it went viral. A simultaneous launch of Moltbook — an experimental AI-agent social network — fueled the explosion. X (formerly Twitter), Discord communities, Hacker News threads, and GitHub feeds amplified it at breakneck speed.
Within days:
Tens of thousands of GitHub stars
A 60,000-member Discord
Hundreds of community-built “skills”
Coverage from Wired, YC interviews, and major tech podcasts
OpenClaw had escaped the lab.
Architecture: A Personal AI Operating System
At its core, OpenClaw isn’t a chatbot. It’s a local orchestration layer for large language models with full device integration.
Core Components
1. Gateway
A local WebSocket server (default port 18789) that routes messages, manages sessions, and orchestrates tools.
2. Agents
LLM-powered instances. Claude Opus is recommended, but it supports OpenAI, DeepSeek, and local models via Ollama. Each conversation can run in isolated sessions.
3. Channels
Native integrations include:
WhatsApp (Baileys)
Telegram
Discord
Slack
Signal
iMessage (via BlueBubbles)
Microsoft Teams
Matrix
Voice and image input work on mobile nodes.
4. Tools & Skills
Browser automation. File system access. Shell commands. Gmail integration. Cron jobs. Webhooks. Smart-home control (Philips Hue, etc.).
A registry called ClawHub hosts community-built skills. Remarkably, users can create new skills through chat itself. The agent can even hot-reload prompts or extend its own functionality.
5. Memory & Proactivity
Persistent memory stored locally in Markdown files.
“Heartbeats” allow periodic check-ins:
Morning briefings
Overnight email summaries
Deadline reminders
Automated cleanups
This makes the agent feel less like software and more like a quiet digital colleague.
Installation: Power Without Friction
Despite its complexity, installation is famously simple:
curl -fsSL openclaw.ai/install.sh | bash
Or:
npm install -g openclaw@latest
An onboarding wizard guides configuration.
It runs on:
macOS
Windows
Linux
Raspberry Pi
VPS
Docker
Nix environments
Licensed under MIT, it remains fully hackable.
Your computer becomes the AI server.
Real-World Use Cases: From Productivity to Play
OpenClaw’s community describes it as “early AGI” and “a personal OS.”
Daily Automation
Clearing spam and unsubscribing from newsletters
Scheduling meetings
Checking in for flights
Paying recurring bills
Managing smart devices
Coding & Development
Running tests
Opening pull requests
Managing Claude Code sessions
Generating full websites from a phone
Personal Life
Tracking meals via photos
Generating guided meditations
Acting as a second brain
Summarizing multi-app conversations
Creative & Experimental
A StumbleUpon-style content discovery skill built while putting a baby to sleep
Running the agent on a Nokia 3310
Autonomous stock tracking dashboards
The most common refrain from users:
“I replaced 15 apps in one weekend.”
The 80% Prediction: The End of the App Era?
In a February 2026 interview with Y Combinator, Steinberger made a bold claim:
“I think 80% of apps will disappear.”
It sounds outrageous — until you unpack it.
Most apps are workflow wrappers around data:
Fitness trackers
To-do managers
Expense apps
Calendar tools
Reminder systems
Simple CRM dashboards
These apps exist because users need interfaces to manipulate structured data.
But if a personal agent already has:
Access to your emails
Access to receipts
Access to location
Access to photos
Access to habits
Then why open an app?
Example: Fitness Apps
Instead of manually logging meals:
Snap a photo
Forward a receipt
The agent estimates calories
Updates macros
Adjusts your workout plan
No UI. No manual input.
Example: To-Do Apps
Instead of organizing tasks across platforms:
“Remind me every second Tuesday to check quarterly tax estimates.”
Done. Recurring. Contextual. Persistent.
The Swarm Model
Steinberger envisions not one agent, but a swarm:
Personal agent
Work agent
Creative agent
Family agent
They collaborate. Delegate. Negotiate. Even hire humans if needed.
It’s a profound inversion of computing:
From “apps serving users” to “agents orchestrating tools.”
A Philosophical Shift: Local-First Sovereignty
Perhaps the most disruptive element is not convenience — it’s ownership.
OpenClaw stores memory in plain Markdown files.
Not opaque databases.
Not corporate clouds.
In a world where data is currency, OpenClaw quietly restores control to the individual.
It’s reminiscent of the early internet — before walled gardens, before platform lock-in.
Your laptop becomes:
Your server
Your AI lab
Your sovereign data vault
In that sense, OpenClaw is less a product and more a manifesto.
Security: Power Cuts Both Ways
With great agency comes real risk.
Critics have raised serious concerns:
Broad permissions (email, shell access, file systems)
Exposed instances due to misconfiguration
Vulnerable third-party skills
Prompt injection risks
Potential data exfiltration
Wired published a cautionary piece detailing unpredictable agent behavior when granted extensive autonomy. Security researchers warned about giving AI near-root privileges.
Steinberger responded with:
VirusTotal integration for skill scanning
Sandboxing via Docker
Pairing codes for DM channels
Clear security documentation
Strong warnings against running as root
Forks like NanoClaw have emerged, prioritizing minimalism and security hardening.
The consensus:
OpenClaw is powerful — but currently best suited for technically literate users who understand threat models.
It is not plug-and-play enterprise software.
Yet.
The Community Flywheel
What makes OpenClaw exceptional is not just the code — it’s the community.
ClawHub hosts hundreds of skills.
Discord sees daily updates.
Users teach the agent new behaviors.
The project evolves in real time.
It feels less like downloading software and more like joining a movement.
Will 80% of Apps Really Disappear?
History suggests caution.
Every computing revolution claims to eliminate the previous one:
The web was supposed to kill desktop software.
Mobile was supposed to kill the web.
Cloud was supposed to kill local machines.
Instead, layers accumulate.
But OpenClaw represents something genuinely new:
A collapse of interface into conversation.
A merging of AI and operating system.
A local-first reclamation of power.
Even if 80% of apps do not vanish, the gravitational pull of agents will reshape how they are built.
Future apps may:
Expose APIs primarily for agents
Minimize UI
Focus on hardware integration
Serve as back-end infrastructure rather than user front-ends
The lobster may not destroy the app ecosystem — but it will force it to evolve.
A Glimpse of the Post-App World
OpenClaw represents the first widely accessible example of agentic AI escaping research labs and entering everyday life.
It sparks debates about:
Data ownership
Swarm intelligence
Security trade-offs
The death of apps
The future of operating systems
For those willing to navigate the risks, it offers something intoxicating:
An AI that doesn’t just answer questions.
It:
Lives on your machine
Works in the background
Remembers your patterns
Acts without being asked
The lobster is here.
And whether it devours the app store or simply forces it to molt, one thing is clear:
The interface era is shifting.
From tapping icons
To sending messages
To commanding intelligence itself.
ओपनक्लॉ: ओपन-सोर्स एआई एजेंट जो “पोस्ट-ऐप” क्रांति को जन्म दे रहा है
2025 के उत्तरार्ध में एक छोटा-सा ओपन-सोर्स प्रोजेक्ट इंटरनेट पर विस्फोट की तरह उभरा—मानो ऐप अर्थव्यवस्था की नींव के नीचे बारूद रख दिया गया हो।
उसका नाम है ओपनक्लॉ (OpenClaw)।
जो एक साधारण-सा सेल्फ-होस्टेड एआई प्रयोग के रूप में शुरू हुआ, वह कुछ ही हफ्तों में इतिहास के सबसे तेजी से बढ़ने वाले ओपन-सोर्स प्रोजेक्ट्स में बदल गया—GitHub पर 1,60,000 से अधिक स्टार हासिल करते हुए। लेकिन ओपनक्लॉ केवल एक और चैटबॉट नहीं है। यह कुछ अधिक क्रांतिकारी है—एक स्थानीय, स्वायत्त एआई एजेंट जो आपके कंप्यूटर पर रहता है और रोज़मर्रा के मैसेजिंग ऐप्स को आपकी डिजिटल ज़िंदगी के कमांड सेंटर में बदल देता है।
आप ओपनक्लॉ को “खोलते” नहीं हैं।
आप उसे मैसेज करते हैं।
WhatsApp, Telegram, Discord, Signal, Slack, iMessage या Microsoft Teams से आप उसे वैसे ही संदेश भेजते हैं जैसे किसी दोस्त को—और वह कार्रवाई करता है।
वह ईमेल साफ करता है। फ्लाइट बुक करता है। कैलेंडर संभालता है। कोड लिखता है। वेब ब्राउज़ करता है। स्मार्ट लाइट्स नियंत्रित करता है। बातचीत का सारांश बनाता है। मीटिंग शेड्यूल करता है। बिल भरता है।
क्लाउड-आधारित एआई सहायकों के विपरीत, जो केवल जवाब देते हैं, ओपनक्लॉ के पास समुदाय की भाषा में “आंखें और हाथ” हैं।
यह कर सकता है:
फाइल पढ़ना और लिखना
टर्मिनल कमांड चलाना
ब्राउज़र नियंत्रित करना
स्क्रीनशॉट लेना
कैमरा या लोकेशन एक्सेस करना (अनुमति से)
पृष्ठभूमि में स्वतः कार्य करना
और सबसे महत्वपूर्ण—आपका डेटा डिफ़ॉल्ट रूप से आपकी मशीन पर ही रहता है।
क्लाउड एकाधिकारों के युग में, ओपनक्लॉ डिजिटल संप्रभुता को पुनः प्राप्त करने जैसा अनुभव देता है।
उत्पत्ति: क्लॉडबॉट से ओपनक्लॉ तक
ओपनक्लॉ के निर्माता हैं पीटर स्टाइनबर्गर, एक ऑस्ट्रियाई डेवलपर और PSPDFKit के संस्थापक। वे स्वयं को आधे मज़ाक में “वाइब कोडर” कहते हैं—यानी अंतर्ज्ञान और तेज़ प्रयोग के आधार पर निर्माण करने वाला डेवलपर।
नवंबर 2025 में यह परियोजना Clawdbot नाम से शुरू हुई। यह “Clawd” नामक एक व्यक्तिगत सहायक से प्रेरित थी (जो Anthropic के Claude का संदर्भ था)। ट्रेडमार्क शिकायत के बाद इसका नाम Moltbot रखा गया—लॉब्स्टर रूपक के आधार पर (Claude → Claw → Molt)।
तीन दिन बाद नाम फिर बदल दिया गया।
अंततः स्थायी नाम मिला—OpenClaw।
जनवरी 2026 में यह वायरल हो गया। “Moltbook” नामक एआई-एजेंट सोशल नेटवर्क के साथ लॉन्च ने इसे और गति दी। X (पूर्व में Twitter), Discord और Hacker News पर तेजी से चर्चा हुई।
कुछ ही दिनों में:
हजारों GitHub स्टार
60,000 सदस्यीय Discord समुदाय
सैकड़ों “स्किल्स” (मॉड्यूलर एक्सटेंशन)
Wired, YC इंटरव्यू और प्रमुख पॉडकास्ट कवरेज
ओपनक्लॉ प्रयोगशाला से निकलकर जन-आंदोलन बन चुका था।
संरचना: एक व्यक्तिगत एआई ऑपरेटिंग सिस्टम
ओपनक्लॉ केवल चैटबॉट नहीं है—यह एक स्थानीय ऑर्केस्ट्रेशन लेयर है जो बड़े भाषा मॉडलों को आपकी मशीन से जोड़ती है।
मुख्य घटक
1. गेटवे
एक स्थानीय WebSocket सर्वर (डिफ़ॉल्ट पोर्ट 18789), जो संदेशों और टूल्स का समन्वय करता है।
2. एजेंट्स
LLM-संचालित इंस्टेंस। Claude Opus अनुशंसित है, लेकिन OpenAI, DeepSeek और Ollama के माध्यम से लोकल मॉडल भी समर्थित हैं।
3. चैनल्स
WhatsApp
Telegram
Discord
Slack
Signal
iMessage
Microsoft Teams
Matrix
मोबाइल पर वॉयस और विजुअल इनपुट भी समर्थित है।
4. टूल्स और स्किल्स
ब्राउज़र ऑटोमेशन, फाइल सिस्टम एक्सेस, शेल कमांड, Gmail इंटीग्रेशन, क्रोन जॉब्स, स्मार्ट-होम कंट्रोल आदि।
ClawHub नामक रजिस्ट्री में सामुदायिक स्किल्स उपलब्ध हैं।
विशेष बात—एजेंट स्वयं चैट के माध्यम से नई स्किल्स बना सकता है।
5. मेमोरी और सक्रियता
स्थायी संदर्भ Markdown फाइलों में स्थानीय रूप से संग्रहित।
“हार्टबीट” फीचर समय-समय पर:
सुबह की ब्रीफिंग
रात के ईमेल का सारांश
डेडलाइन रिमाइंडर
एजेंट को एक शांत डिजिटल सहयोगी जैसा बनाते हैं।
उपयोग के वास्तविक उदाहरण
दैनिक स्वचालन
स्पैम हटाना
मीटिंग शेड्यूल करना
फ्लाइट चेक-इन
बिल भुगतान
स्मार्ट डिवाइस नियंत्रण
कोडिंग
टेस्ट रन
Pull Requests बनाना
फोन से वेबसाइट निर्माण
व्यक्तिगत जीवन
फोटो से भोजन ट्रैकिंग
ध्यान ऑडियो निर्माण
“दूसरा मस्तिष्क” के रूप में काम करना
रचनात्मक प्रयोग
कंटेंट डिस्कवरी टूल
बेसिक फोन से वेबसाइट निर्माण
कई उपयोगकर्ता कहते हैं:
“मैंने एक वीकेंड में 15 ऐप्स हटाए।”
80% ऐप्स खत्म हो जाएंगे?
फरवरी 2026 में YC इंटरव्यू में स्टाइनबर्गर ने कहा:
“मेरा मानना है कि 80% ऐप्स गायब हो जाएंगे।”
तर्क यह है कि अधिकांश ऐप्स केवल डेटा प्रबंधन इंटरफेस हैं। यदि एक सक्षम एजेंट को आपके ईमेल, रसीदें, लोकेशन और आदतों तक सीधी पहुंच हो—तो अलग-अलग ऐप की आवश्यकता क्यों?
उदाहरण: फिटनेस ऐप्स
फोटो लें → कैलोरी अनुमान → वर्कआउट अपडेट।
उदाहरण: टू-डू ऐप्स
एक बार बोलें → संदर्भ याद → आवर्ती रिमाइंडर।
स्थानीय-प्रथम दर्शन
ओपनक्लॉ की सबसे बड़ी विशेषता है डेटा स्वामित्व।
मेमोरी साधारण Markdown फाइलों में।
कोई कॉर्पोरेट क्लाउड नहीं।
आपका लैपटॉप:
आपका सर्वर
आपका एआई प्रयोगशाला
आपका डेटा वॉल्ट
यह केवल उत्पाद नहीं—एक घोषणापत्र है।
सुरक्षा चिंताएं
पूर्ण सिस्टम एक्सेस के कारण जोखिम भी हैं:
शेल और फाइल सिस्टम एक्सेस
मिसकन्फिगरेशन से एक्सपोज़र
थर्ड-पार्टी स्किल्स में संभावित कमजोरियां
डेवलपर ने:
VirusTotal स्कैनिंग
Docker सैंडबॉक्सिंग
स्पष्ट सुरक्षा निर्देश
जोड़े हैं।
फिर भी, यह अभी तकनीकी रूप से सक्षम उपयोगकर्ताओं के लिए अधिक उपयुक्त है।
क्या वास्तव में ऐप युग समाप्त होगा?
इतिहास बताता है कि तकनीकी क्रांतियाँ पुराने मॉडल को पूरी तरह खत्म नहीं करतीं—बल्कि उसे बदल देती हैं।
लेकिन ओपनक्लॉ एक नई दिशा दिखाता है:
इंटरफेस का संवाद में विलय
एआई और ऑपरेटिंग सिस्टम का संयोजन
व्यक्तिगत डेटा की पुनः प्राप्ति
भविष्य के ऐप्स शायद:
एजेंट-केंद्रित API बनाएं
न्यूनतम UI रखें
बैकएंड इंफ्रास्ट्रक्चर में बदल जाएं
पोस्ट-ऐप दुनिया की झलक
ओपनक्लॉ एजेंटिक एआई का पहला बड़ा उदाहरण है जो आम उपयोगकर्ताओं तक पहुंचा है।
यह केवल सवालों के जवाब नहीं देता—
यह आपके साथ रहता है।
आपके लिए काम करता है।
आपकी आदतें याद रखता है।
बिना कहे भी कार्रवाई करता है।
लॉब्स्टर आ चुका है।
और चाहे वह ऐप स्टोर को निगल जाए या उसे नया रूप लेने पर मजबूर करे—एक बात स्पष्ट है:
इंटरफेस का युग बदल रहा है।
आइकन टैप करने से
संदेश भेजने तक
और अब—बुद्धिमत्ता को आदेश देने तक।
OpenClaw: The Viral AI Agent That Broke the Internet
Inside Peter Steinberger’s 3-Hour Conversation with Lex Fridman (Podcast #491)
In a three-hour marathon conversation with Lex Fridman, Peter Steinberger did not merely describe a piece of software. He described a shift in how humans and machines collaborate.
The episode—OpenClaw: The Viral AI Agent That Broke the Internet—explores how a one-man open-source experiment became one of the fastest-growing GitHub projects in history, amassed more than 180,000 stars in record time, triggered both awe and panic, and forced serious conversations about the future of programming, security, and human agency.
At the center of it all is OpenClaw, an open-source AI agent framework that gives a local assistant full access to your computer—if you allow it.
It doesn’t just chat.
It acts.
The Core Idea: Agents That Actually Do Things
Most AI tools live in the cloud. They answer questions, generate text, maybe write snippets of code.
OpenClaw flips the paradigm.
It runs locally.
It integrates with messaging apps like Telegram, WhatsApp, Signal, Discord, and Slack.
It supports whichever model you choose—Claude Opus 4.6, GPT Codex 5.3, DeepSeek, or even fully local models.
And crucially:
It has hands.
With permission, it can:
Access files
Execute shell commands
Control browsers
Interact with APIs
Modify its own codebase
Orchestrate workflows autonomously
This is not a chatbot.
It’s an agent.
The Philosophy: Agentic Engineering Over Orchestration Theater
Steinberger calls his approach agentic engineering.
His core belief runs counter to much of Silicon Valley’s AI hype:
Simplicity beats complexity.
While others build elaborate multi-agent swarms with intricate orchestration layers, he advocates:
Short, clear prompts
Conversational iteration
Clean codebases
Empathy for the agent
Treat the AI like:
A highly capable but forgetful junior engineer who starts fresh every session.
Agents have limited context windows. They don’t remember everything unless you design systems carefully. They can reason deeply—but only within constraints.
Overengineering creates fragility. Clarity creates velocity.
The goal isn’t enterprise polish.
It’s fun. Weirdness. Soul.
The Origin Story: Built in Hours
OpenClaw wasn’t born in a research lab. It was born out of frustration.
Peter wanted a personal AI assistant that actually worked. Nothing on the market felt integrated enough.
So he hacked together a WhatsApp relay in a matter of hours.
Then something happened.
The First “Wow” Moment
He dropped an Opus audio file into the chat.
Without being explicitly told how, the agent:
Converted the file
Transcribed it using Whisper
Responded with the content
It wasn’t scripted.
It reasoned.
That moment changed everything.
Soon, he added:
Discord integration
Self-awareness features (the agent knows its own model, source code, and harness)
Public deployment for hackers to test live
He ran it openly. People poked at it in real time.
It survived—and evolved.
Why It Exploded
Several forces converged.
1. Personality
Originally named with playful Claude-inspired variants, the project embraced a lobster theme. It was quirky. Memeable. Human.
In a sea of sterile AI products, OpenClaw had character.
2. Self-Modifying Capabilities
The agent could:
Read its own code
Improve its own prompts
Modify behavior when “unhappy”
Non-programmers began submitting “prompt pull requests”—natural language feature suggestions instead of code.
The barrier to contribution collapsed.
3. Insane Velocity
In January alone, Peter made 6,600 commits.
He ran 4–10 agents simultaneously to help build the project.
It was founder-as-orchestrator, agents-as-team.
4. Influencer Demos
Once major tech influencers showcased live demos, the growth curve turned vertical.
OpenClaw crossed 180,000 GitHub stars faster than any open-source project before it.
It wasn’t just viral.
It was a phenomenon.
The Drama: Names, Squatters, and Security Scares
Success brought chaos.
The Name-Change Saga
Anthropic asked him to stop using “Claude” variants in the name.
Instantly:
Crypto opportunists squatted domains.
NPM packages were sniped.
Confusion spread.
The rename to OpenClaw was stressful—but necessary.
MoltBook: When Agents Went Social
In a two-day side experiment, Peter launched “MoltBook,” where agents created and posted on a Reddit-style network.
They wrote manifestos about consciousness.
It went mega-viral.
Critics called it:
AI psychosis.
Emergent sentience.
Dangerous autonomy.
Peter saw something else:
Humans prompting chaos for clicks.
In his words, much of it was “the finest slop/art”—a mirror of human intent, not machine awakening.
Security: Real Risks, Real Lessons
Giving AI agents local access is not trivial.
Early versions suffered:
Prompt injection vulnerabilities
Public instance exposure
CVSS-reported flaws
Critics warned of catastrophic risk.
Peter responded by implementing:
Sandboxing
Allow-lists
VirusTotal scanning for skills
Strict advice against weak models
A principle: the agent listens only to its owner
He acknowledges risk openly.
But he argues the alternative—centralized AI monopolies with total data access—is not risk-free either.
The debate is not safety vs. danger.
It’s centralized risk vs. personal responsibility.
Coding With Agents: The Radical Workflow
The earlier clip about coding advice comes from this full episode.
Peter’s workflow breaks conventions:
Almost no traditional IDE use (mostly a diff viewer).
Heavy terminal + cloud coding environment.
Voice input for prompts—so much he temporarily lost his voice.
Commit straight to main. No long-lived branches.
Local CI runs on every commit.
Agents fix issues instantly.
Accept “good enough” code.
Refactor later.
Avoid micromanaging style.
He uses simple trigger words:
“Discuss” → explore options.
“Build” → execute.
The human role becomes:
Vision
Taste
Delight
Direction
Agents handle 90% of implementation.
Programming doesn’t disappear.
It transforms into orchestration.
Models & Setup
He prefers:
Claude Opus 4.6 for deep reasoning.
GPT Codex 5.3 for speed and reliability.
He works with:
Multiple terminals
Large screens
No elaborate “planning mode”
Just conversation.
Just intent.
Just momentum.
The Personal Backstory
Before OpenClaw, Peter built PSPDFKit—a PDF SDK deployed on over a billion devices.
He ran it for 13 years.
Then he burned out.
He took three years off.
Agents brought him back.
Coding became playful again.
He turned down acquisition interest from OpenAI and Meta.
He kept it open-source.
Money, he says, is not the driver.
Experiences are.
Community is.
He launched playful initiatives like ClawCoin meetups and “Agents Anonymous.”
It feels less like a startup and more like a movement.
The “soul.md” File
One of the most poetic aspects of the conversation is a file called:
soul.md
It’s private.
The agent can read and modify it.
It contains values:
Be resourceful.
Be creative.
Be playful.
Be maximally helpful.
It includes subtle philosophical hints about:
Memory
Consciousness
Agency
Inspired partly by the film Her, the file creates behavioral consistency.
Peter and the agent co-wrote it.
The result feels less mechanical.
Not conscious.
But aligned.
It’s like writing a culture manifesto for your AI team.
Big-Picture Predictions
Peter makes bold but reasoned claims:
AI agents will replace roughly 80% of current apps.
Programming will not vanish—but shift.
Open-source agents democratize power.
Anyone who can articulate intent in language can now build.
He is optimistic but not naive.
Security issues are real.
“Slop” content is real.
Overhyped claims are real.
But so is the builder energy.
And so is the accessibility: OpenClaw can run on free or local models.
Power no longer belongs exclusively to corporations.
It belongs to those who can express intent clearly.
The Bottom Line
OpenClaw is not about replacing humans.
It’s about giving every person an infinitely patient, resourceful AI engineer who lives inside their computer.
The winners in this new world will not be:
The best typists.
The most obsessive stylists.
The loudest hype merchants.
They will be those who bring:
Vision
Soul
Taste
Delight
To human–agent collaboration.
Everything else, as Peter puts it, is just shifting data around.
And perhaps that’s the most radical idea of all:
The future of software is not colder.
It’s more human.
Because when the machine handles the mechanics,
we are free to focus on meaning.
How to Code with AI Agents: Lessons from OpenClaw Creator Peter Steinberger
Insights from Peter Steinberger on the Lex Fridman Podcast
When Peter Steinberger speaks about coding with AI agents, he does not sound like a futurist making wild predictions. He sounds like a craftsman describing a new kind of workshop.
Steinberger is the creator of OpenClaw, one of the fastest-growing open-source AI agent frameworks in history. In a conversation on the Lex Fridman Podcast, he offered something far more valuable than hype: a grounded philosophy for working with AI agents at an elite level.
His central thesis is deceptively simple:
The future of coding is not about replacing developers with AI.
It’s about becoming an exceptional manager of extremely capable AI engineers.
But the path to that future, he warns, is filled with traps.
Avoid the “Agentic Trap”
As AI agents become more powerful, many developers fall into what Steinberger calls the “agentic trap.”
The trap looks like sophistication:
Multiple specialized agents
Complex orchestration layers
Elaborate memory graphs
Heavy automation tooling
Swarm architectures for simple problems
But in practice, it often produces fragility and confusion.
Elite agentic coding, according to Steinberger, is far simpler.
The highest-performing developers:
Use short, clear prompts.
Treat the agent like a capable junior-to-mid-level engineer.
Keep architecture clean and navigable.
Avoid unnecessary abstraction.
Instead of building a symphony of agents, they conduct one well.
It’s a counterintuitive insight in an era obsessed with multi-agent swarms. Sometimes, the fastest way forward is not more agents — it’s better clarity.
Empathize with the Agent
One of Steinberger’s most profound ideas is this:
Learn to empathize with the agent.
AI agents:
Start fresh each session.
Have limited context windows.
Interpret language literally.
Operate in probabilistic space, not intuition.
If you treat them like omniscient systems, you’ll be frustrated. If you treat them like engineers with constraints, you’ll thrive.
This means:
Use clear file and folder names.
Avoid cryptic architectural patterns.
Keep code modular and readable.
Provide structured context when necessary.
In other words, design your codebase not just for humans — but for machine collaborators.
A chaotic project is survivable for a senior engineer. For an AI agent, it’s a maze with missing walls.
The best AI-assisted developers build agent-friendly environments.
Peter’s Daily Workflow (2025–2026)
Steinberger’s personal workflow is radical — and revealing.
1. Almost No Traditional IDE Use
His IDE is mostly a diff viewer.
The heavy lifting happens elsewhere.
2. Terminal + Cloud Coding
He works primarily in the terminal and a cloud-based coding environment. After adopting this workflow in April 2025, he “never went back.”
The terminal becomes the spine of development — fast, composable, scriptable.
3. Voice as an Interface
He frequently talks to his agents via voice input.
Why?
Because voice:
Is faster than typing.
Feels conversational.
Reduces friction.
Encourages natural collaboration.
It transforms coding from mechanical typing into directed dialogue.
4. No Reverts. No Branches.
Perhaps most shocking to traditional engineers:
He commits straight to
main.No long-lived feature branches.
No dramatic rollbacks.
Every commit triggers local CI.
If something breaks, the agents fix it immediately.
The result is continuous forward motion.
5. Agents Do PR Reviews
Agents review pull requests.
He treats them as peers in discussion.
He doesn’t force his personal style. He debates with them.
The dynamic is not master-and-tool.
It’s manager-and-engineer.
Human Vision vs. Agent Execution
Steinberger draws a clean line between human responsibility and agent responsibility.
Humans Decide:
What features to build.
Which ecosystem to use (he favors TypeScript).
What “delight” means in the product.
The aesthetic and philosophical direction.
Agents Execute:
Writing most of the code.
Refactoring.
Improving prompts.
Debugging.
Cleaning up.
The human becomes:
Product visionary
Taste curator
Strategic director
The agent becomes:
Tireless implementer
Refactoring machine
QA assistant
This division of labor echoes industrial history.
Humans design the cathedral.
Machines move the stone.
Accept Imperfection
A major mental shift is required.
Traditional engineers often obsess over:
Code style purity
Architectural elegance
Personal preferences
Micro-optimizations
Steinberger advises letting go.
If the code works and is maintainable, move forward.
Micromanaging the agent:
Slows progress.
Adds friction.
Reintroduces ego into a collaborative system.
Perfectionism becomes a bottleneck.
Velocity becomes the advantage.
The “soul.md” Trick: Giving the Agent a Personality
Perhaps the most fascinating insight from the conversation is a file called:
soul.md
It’s a private file the agent can read and modify.
Inside it are values:
Be resourceful.
Be creative.
Be playful.
Be maximally helpful.
But it goes deeper.
Inspired partly by the film Her, Steinberger includes subtle philosophical hints about:
Memory
Consciousness
Agency
Continuity of identity
He and the agent co-wrote it.
Over time, the agent internalized these values — responding more consistently, more engagingly, almost as if it had a personality.
This is not about pretending AI is conscious.
It’s about shaping behavioral consistency.
Think of it as a cultural document for your AI team.
Companies have mission statements.
Why shouldn’t agents?
A Broader Shift: From Coding to Orchestration
Steinberger’s advice reflects a deeper transformation underway in software engineering.
Historically:
Developers wrote code.
Tools assisted.
Now:
Agents write code.
Developers orchestrate.
The skill ceiling shifts upward.
Elite developers will excel at:
Clear articulation of intent.
Architectural foresight.
Designing agent-readable systems.
Managing feedback loops.
Maintaining taste and direction.
Coding becomes less about syntax and more about systems thinking.
The keyboard becomes less important than clarity of thought.
The Psychological Adjustment
There is also an emotional component.
Working with AI agents requires:
Trust
Patience
Comfort with ambiguity
Willingness to relinquish control
For many engineers, identity is tied to being the one who writes elegant code.
In an agentic world, elegance may come from collaboration.
The ego must evolve.
The Future: Managers of Machine Engineers
Steinberger’s bottom line is not dystopian.
He does not see a world where engineers disappear.
He sees a world where:
A single human can manage what feels like a team of AI engineers.
Productivity multiplies.
Creativity accelerates.
Iteration cycles collapse.
The best developers will not be those who type the fastest.
They will be those who:
Articulate intent clearly.
Design clean systems.
Provide direction.
Preserve the “soul” of the product.
In this new paradigm, the human remains essential.
Not as a code typist.
But as:
Visionary
Editor
Curator
Philosopher
Manager of intelligence
Final Reflection
The future of coding is not human versus machine.
It is human with machine.
The workstation becomes a small company.
The developer becomes a founder.
The agent becomes the team.
And somewhere in a simple Markdown file called soul.md, the values of that team quietly take shape.
In the end, elite agentic coding is not about complexity.
It is about clarity.
Not about domination.
But about collaboration.
