#491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger
In this episode of the Lex Fridman Podcast, Peter Steinberger discusses the development of OpenClaw, an open-source AI project that started as a simple WhatsApp-to-Cloud Code prototype and evolved into a social network called Moldbook. He explains the technical architecture behind OpenClaw's AI agent system, including its features for natural interactions and integration with various messaging platforms.
Steinberger shares his perspective on how AI agents could reshape the technology landscape, suggesting that traditional apps might become less relevant as AI gains the ability to execute tasks directly. He also addresses the future role of human developers, noting that while AI can now modify its own software, human creativity and problem-solving skills remain important as programming evolves toward more design-focused work.
This is a preview of the Shortform summary of the Feb 12, 2026 episode of the Lex Fridman Podcast
Sign up for Shortform to access the whole episode summary along with additional materials like counterarguments and context.
1-Page Summary
The Development and Evolution of the Openclaw Project
Peter Steinberger shares the journey of Openclaw, an open-source AI project that began as a simple one-hour prototype connecting WhatsApp to Cloud Code through CLI. Despite facing significant challenges, including online harassment and naming rights issues, Steinberger successfully transformed the project into a social network called Moldbook, focusing on making AI technology accessible to both technical and non-technical users.
Technical Details of Openclaw's Agentic AI Architecture
Steinberger explains that Openclaw's architecture revolves around an agentic loop that makes AI interactions more natural by incorporating features like selective responses and smart queuing. The system includes various components such as gateways and chat clients, allowing seamless integration with platforms like WhatsApp, Telegram, and Discord. Steinberger notes that the project has evolved to prefer CLI-based skill integration, making it easier for the AI to understand and implement new capabilities.
Agentic AI's Impact on Software, Apps, and the Tech Landscape
According to Steinberger, Agentic AI is poised to revolutionize how users interact with technology. He predicts that many traditional apps may become obsolete as AI agents gain the ability to execute tasks directly without specialized software intermediaries. While emphasizing the importance of security and user privacy, Steinberger envisions a future where AI's predictive and agentic qualities create a more seamlessly integrated tech experience.
Future of Programming and Human Developers in AI World
While Steinberger acknowledges that AI can now modify its own software, he emphasizes that human creativity and problem-solving skills remain essential. He suggests that programming may evolve into a more creative pursuit, with developers focusing on design and complex problem-solving rather than just coding. Steinberger encourages developers to view themselves as builders who can adapt their skills to work alongside AI, contributing unique insights to user experience and system integration.
1-Page Summary
Additional Materials
Clarifications
- Agentic AI refers to artificial intelligence systems designed to act autonomously, making decisions and taking actions on behalf of users. An agentic loop is a feedback cycle where the AI continuously perceives its environment, decides on actions, executes them, and learns from the outcomes to improve future behavior. This loop enables more dynamic, context-aware, and goal-directed interactions compared to static AI responses. The significance lies in creating AI that can proactively manage tasks and adapt without constant human input.
- CLI, or Command Line Interface, is a text-based way to interact with software by typing commands. It is preferred for skill integration because it allows precise, scriptable control over AI functions without the complexity of graphical interfaces. This makes it easier to automate tasks and extend AI capabilities efficiently. Additionally, CLI tools are lightweight and can be used across different systems consistently.
- Selective responses allow an AI to choose the most relevant or appropriate reply from multiple options based on context or priority. Smart queuing manages incoming tasks or messages by organizing and prioritizing them to avoid overload and ensure timely handling. Together, these features make AI interactions feel more natural and efficient by mimicking human-like attention and decision-making. This approach helps the AI focus on important inputs while delaying or ignoring less critical ones.
- Gateways act as intermediaries that translate messages between the AI system and external platforms, handling protocol differences and authentication. Chat clients are software interfaces that manage conversations, sending user inputs to the AI and displaying its responses. Together, they enable the AI to communicate seamlessly across different messaging services by adapting to each platform's unique requirements. This setup allows the AI to function as a user or bot within apps like WhatsApp, Telegram, and Discord.
- Naming rights issues refer to legal or trademark conflicts that arise when a project’s name is similar or identical to an existing brand or product. These issues can force developers to change the project’s name to avoid infringement claims. Resolving naming rights often involves negotiations or rebranding to ensure the name is unique and legally safe. This process can delay development and affect the project's public identity.
- AI modifying its own software involves algorithms that can analyze and change their code or parameters to improve performance. This process often uses techniques like reinforcement learning or evolutionary algorithms to test and implement changes autonomously. Such self-modification enables AI to adapt to new tasks or environments without human intervention. However, safeguards are necessary to prevent unintended or harmful alterations.
- Traditional apps are designed for specific tasks and require users to interact with distinct interfaces. Agentic AI can perform multiple tasks autonomously by understanding user intent and context, reducing the need for separate apps. This shift could simplify user experience by consolidating functions into AI agents rather than individual applications. It also challenges current software business models that rely on app distribution and usage.
- Predictive capabilities refer to AI's ability to analyze data and forecast outcomes based on patterns. Agentic qualities involve AI taking autonomous actions to achieve goals, not just making predictions. Prediction is passive, while agency implies active decision-making and task execution. Agentic AI can initiate and manage tasks independently, beyond simply suggesting possibilities.
- Programming evolving into a "more creative pursuit" means developers will spend less time writing routine code and more time designing innovative solutions. AI can handle repetitive coding tasks, freeing humans to focus on complex challenges and user experience. This shift emphasizes strategic thinking, architecture, and problem-solving over manual coding. Creativity in programming will involve imagining new possibilities and guiding AI tools effectively.
- Human developers guide AI by designing frameworks and setting goals that shape AI behavior. They ensure AI systems align with user needs and ethical standards. Developers also handle complex integration tasks that require nuanced understanding beyond AI's current capabilities. This collaboration enhances overall user experience by combining human creativity with AI efficiency.
Counterarguments
- While Openclaw's agentic loop may improve AI interactions, there is a risk that over-reliance on AI could lead to a loss of human touch in communication and a decrease in personal skills in managing and interpreting interactions.
- The integration of AI across various platforms may raise concerns about platform dependency and monopolization, where a few large tech companies could control the AI ecosystem.
- The claim that agentic AI will render many traditional apps obsolete may be premature, as there will likely always be specialized tasks and niche markets that require dedicated software solutions.
- Emphasizing security and user privacy is important, but there may be skepticism about whether these considerations can keep pace with the rapid development of AI technology and its integration into social networks like Moldbook.
- The idea that AI's predictive and agentic features will create a more seamlessly integrated technology experience assumes that all users will be comfortable with, and have access to, such advanced technology, which may not be the case due to digital divides.
- The assertion that human creativity and problem-solving remain essential in programming may overlook the potential for AI to significantly outpace human capabilities in certain areas of creativity and complex problem-solving in the future.
- Suggesting that programming will evolve into a more creative discipline may not account for the possibility that AI could automate many aspects of design and problem-solving, potentially reducing the need for human input in these areas.
- Encouraging developers to adapt their skills to work alongside AI assumes that all developers have the capacity and resources to do so, which may not be the case for everyone, especially in under-resourced or marginalized communities.
Get access to the context and additional materials
The Development and Evolution of the Openclaw Project
Peter Steinberger shares the story of Openclaw's evolution from a one-hour prototype to a powerful open-source AI project that aims to be fun, engaging, and accessible for all users.
Peter Steinberger's Journey In Building Openclaw, a Successful Open-Source Ai Project
Steinberger's journey in developing Openclaw began with a revolutionary one-hour prototype and evolved through various challenges, including harassment and naming rights.
Steinberger's One-hour Prototype Ignited Openclaw's Growth
Peter Steinberger built a transformative prototype in just one hour by connecting WhatsApp to Cloud Code through the command line interface (CLI). This allowed Steinberger to communicate with his computer via chat, and while the initial build was limited, it had powerful capabilities.
Challenges Steinberger Faced Developing and Naming Openclaw, Including "Waa Relay" and "Moldbot" Saga
Steinberger had to maneuver through substantial hurdles, not only technical but also social and legal ones. He faced online harassment from the cryptocurrency community, which stormed his Discord server. To combat disruptions like spamming, Steinberger implemented strict server rules.
Renaming the project added to the adversity. After the first round of name changes, due to the Anthropic request, came a stressful, sleepless marathon to secure domain names, resulting in the temporary and unsatisfying choice of "Moldbot."
He recounted the nerve-wracking experience of losing his account name within seconds to a squatter during the renaming process, highlighting the unpredictability of the situation and the lengths malicious actors would go to disrupt his work.
Security problems emerged for Openclaw, challenging Steinberger with issues like exploits and prompt injection—an industry-wide problem. With clear documentation against certain configurations, Steinberger continues the ongoing battle for security within Openclaw.
Steinberger's Aim: Make Openclaw Fun, Engaging, and Accessible for All, Including Non-technical Users
Despite the trials, S ...
Here’s what you’ll find in our full summary
The Development and Evolution of the Openclaw Project
Additional Materials
Actionables
- You can explore the potential of chatbots by setting up a simple bot on a platform like Telegram or Slack to automate a daily task, such as sending you weather updates or reminders for meetings. Start by using bot creation tools that require minimal coding, like ManyChat or Chatfuel, to get a feel for how bots can simplify your life.
- Enhance your online community's experience by establishing a code of conduct for any group or forum you manage, ensuring a respectful and harassment-free environment. Use free online templates to draft your guidelines and actively moderate the space to foster a positive community culture.
- Broaden your understanding of AI's accessibility by using voice-to-text features on your ...
Get access to the context and additional materials
Technical Details of Openclaw's Agentic AI Architecture
Steinberger and Fridman delve into the intricacies of Openclaw's artificial intelligence setup, discussing everything from its agentic loop and gateway to its innovative applications in natural interactions and problem-solving.
Key Components of Openclaw: Gateway, Chat Clients, Harness, Agentic Loop
Steinberger on Agentic Loop's Role in AI Self-Improvement
Steinberger discusses the critical role the agentic loop plays in Openclaw's AI, making the agent appear more human by sometimes choosing not to reply to create a more natural interaction. He indicates that this loop includes gateways and a harness. The components, such as the gateway, chat interface, and chat applications like WhatsApp, Telegram, Discord, allow Openclaw to interact with the user meaningfully.
Steinberger emphasizes the significance of the agentic loop and smart queuing, enhancing the AI to be more relatable. He also touches on continuous reinforcement learning as an end goal, with current levels utilizing markdown files and databases. He cites the use of a "no reply token" and a "heartbeat" feature that performs actions in a context-driven manner, like checking up on a user post-surgery, to advance the AI's relationship-building and self-improvement attributes.
Extending Openclaw Agent Skills and Tools
Steinberger showcases a passion for growth within Openclaw. He describes an energetic browser user made possible by the project, which integrates Playwright with extras to streamline operations for agents. He underscores his meaningful engagement with Openclaw, gaining attention from people who wish to understand open-source development.
Addressing the expansion of agent skills and tools, he notes collaborating with VirusTotal to vet every skill and identify bugs preemptively. The skills in OpenClaw are described as easily integratable, with a single explanatory sentence enabling the AI to comprehend and apply the related command-line interface (CLI). Most skills work satisfactorily with this method.
Steinberger encounters limitations with async image loading on Apple platforms, but he enthusiastically recounts the project's evolution, where expectations of command pattern interfaces have transitioned to a preference for CLI. Through this, the model can augment its capabilities.
Steinberger's Method For Voice ...
Here’s what you’ll find in our full summary
Technical Details of Openclaw's Agentic AI Architecture
Additional Materials
Actionables
- You can enhance your daily communication by adopting the "no reply" approach in digital conversations to make interactions feel more natural. Instead of feeling compelled to respond immediately to every message, allow yourself to sometimes not reply when it doesn't add value to the conversation. This mimics the agentic loop's behavior and can lead to more meaningful exchanges when you do choose to engage.
- Experiment with using voice commands on your existing devices to streamline complex tasks. For example, if you're cooking and your hands are busy, use voice commands to set timers, play music, or ask for recipe conversions. This practice can help you become more comfortable with natural language interactions and appreciate the convenience of voice over typing, similar to how Openclaw prioritizes voice commands.
- Try using markdown files to organize your pe ...
Get access to the context and additional materials
Agentic AI's Impact on Software, Apps, and the Tech Landscape
Agentic AI is ushering in a transformative era for the technology ecosystem, potentially rendering traditional software and services obsolete through enhancements in task efficiency, while also raising concerns about privacy and security.
Agentic AI's Potential to Disrupt Traditional Software and Services
Lex Fridman and Peter Steinberger discuss how Agentic AI, particularly through platforms like Openclaw, is not just for programmers but is enriching people's lives beyond the tech community. Steinberger puts forth an enthusiastic "builder vibe," insinuating that Agentic AI facilitates creativity and could transform how users interact with software and applications.
Secure, Transparent AI Ensuring User Privacy and Data Control
Steinberger highlights the importance of security in AI applications. He stresses the ongoing focus on implementing robust security features in Agentic AI projects, like Openclaw, to make them safe for users. Addressing security practices and audits, he emphasizes the importance of ensuring user privacy and control over data. This focus reflects a broader emphasis on maintaining trust in the emerging AI-led landscape.
Steinberger also notes the encroaching obsolescence of certain apps, as agents gain the ability to execute tasks without the need for specialized software. He predicts a future where apps have to become APIs, whether they want to or not, because agents can figure out how to utilize a phone's services directly. He uses the example of a Sonos app to illustrate that an agent could communicate with speakers without the traditional intermediary app, indicating a shift towards integrated command execution.
He also hints at a deeper integration of AI into our everyday devices and services. Steinberger suspects that a significant number of existing apps could become obsolete as personal agents may know more about the user and thus make better decisions. Rather than using separate fitness apps, AI could understand a user's physical context and offer tailored workout or dietary suggestions.
Furthermore, Steinberger envisions a more fluid technology ecosystem, where AI's predictive and agentic qualities may negate the necessity to open multiple apps ...
Here’s what you’ll find in our full summary
Agentic AI's Impact on Software, Apps, and the Tech Landscape
Additional Materials
Clarifications
- Agentic AI refers to artificial intelligence systems that can autonomously make decisions and take actions to achieve specific goals without constant human input. Unlike traditional AI, which often requires direct commands or operates within narrow tasks, agentic AI acts proactively and adapts to changing environments. It functions like a digital agent that can plan, learn, and execute complex sequences of tasks independently. This autonomy enables it to interact with multiple software and services seamlessly, transforming user experiences.
- An API (Application Programming Interface) allows different software systems to communicate and share data or functions without a traditional user interface. When apps become APIs, they no longer need to be directly used by people through screens but instead provide services that AI agents can access and control behind the scenes. This means users interact with AI agents that handle multiple app functions seamlessly, rather than opening each app individually. It shifts software usage from manual operation to automated, integrated task execution.
- AI agents can communicate directly with hardware or services by using APIs (Application Programming Interfaces) that expose device functions. These APIs allow software to send commands and receive data without needing a traditional user interface. Advanced AI agents can interpret user instructions and interact with these APIs autonomously. This bypasses the need for separate apps acting as intermediaries.
- Agentic qualities in AI refer to the ability of the AI to act autonomously, make decisions, and perform tasks on behalf of the user. Predictive qualities involve the AI analyzing data to forecast outcomes or suggest actions based on patterns. While predictive AI anticipates needs, agentic AI takes initiative and executes tasks independently. This shift enables AI to move from passive assistance to active problem-solving.
- AI agents gain deeper knowledge of users by continuously collecting and analyzing data from user interactions, preferences, and behaviors across devices and apps. They use machine learning models to identify patterns and predict needs, enabling personalized decision-making. This process often involves integrating data from sensors, calendars, location services, and past activities to build a comprehensive user profile. Privacy-preserving techniques like data anonymization and local processing help protect user information during this learning.
- When apps become obsolete, users no longer need to open multiple separate programs to complete tasks, as AI agents handle these tasks directly. For developers, this means shifting focus from creating standalone apps to building APIs and services that AI agents can access and integrate seamlessly. This transition could reduce app downloads and change revenue models, emphasizing backend functionality over user interfaces. Ultimately, it demands new skills and strategies for developers to remain relevant in an AI-driven ecosystem.
- Security practices in AI development involve designing systems to protect against unauthorized access, data breaches, and manipulation. Audits are systematic evaluations of AI systems to verify compliance with security standards and identify vulnerabilities. These processes include code reviews, penetration testing, and monitoring data handling to ensure privacy and integrity. Together, they help build trustworthy AI that safeguards user information and operates reliably.
- AI technologies often require access to large amounts of personal data to function effectively, raising concerns about how this data is collected, stored, and used. Privacy challenges include risks of unauthorized data access, misuse, and potential surveillance. Debates focus on creating transpa ...
Counterarguments
- Agentic AI may not render traditional software obsolete but instead create a complementary relationship where both coexist and enhance each other's capabilities.
- The enrichment of lives beyond the tech community assumes widespread accessibility and adoption, which may not be uniform due to digital divides and varying levels of tech literacy.
- The shift from traditional interfaces to integrated command execution might not be universally preferred due to diverse user preferences and the potential loss of control or understanding over complex processes.
- Prioritizing robust security features does not guarantee safety, as new technologies often introduce unforeseen vulnerabilities and challenges.
- Ensuring user privacy and control over data is difficult in practice, and there may be significant challenges in maintaining these ideals as AI systems become more complex and integrated into various aspects of life.
- Trust in the AI-led landscape is not solely dependent on transparent security practices and audits but also on the ethical use of AI and the intentions of those who control these technologies.
- The ability of AI agents to execute tasks directly could lead to over-reliance on AI, reducing users' skills and ability to perform tasks independently.
- The necessity for traditional apps to function as APIs may not be absolute, as niche or specialized applications could retain their relevance due to unique features or user preferences.
- Direct communication between AI agents and hardware or services assumes a level of standardization and interoperability that may not be feasible across all devices and platforms.
- The prediction of app obsolescence underestimates the adaptability of app developers and the potential for apps to evolve alongside AI technologies.
- Personalized recommendations from AI based on physical context may not always be accurate or desired, and there cou ...
Get access to the context and additional materials
Future of Programming and Human Developers in AI World
Peter Steinberger discusses the transformation of the programming landscape as AI becomes more integrated into development workflows.
Human Programmers' Role as AI Advances
Steinberger focuses on the ongoing excitement surrounding Openclaw and its empowerment of non-programmers and programmers alike, noting that AI offers new possibilities for human creativity and problem-solving.
Steinberger On Agentic AI Replacing Traditional Programming, Preserving Human Creativity and Problem-Solving
Steinberger observes that Openclaw AI is now capable of modifying its own software. This advancement suggests a decrease in traditional programming tasks but emphasizes that human insight, or the "style, love, that human touch," remains indispensable. Programming may become akin to knitting, something done for the joy of it, while the essence of being a builder, defined by creativity and problem-solving, is preserved.
Steinberger urges developers to view themselves not merely as iOS developers or in other specific roles but as builders who apply their skills in diverse, perhaps unconventional, ways. Developers are encouraged to embrace AI, focusing more on design, overarching problem-solving, and directing AI rather than just coding.
Role of Developers and Engineers in Leveraging Agentic AI for Innovation
Steinberger notes that while AI could replace aspects of programming, the creative process of deciding what to build and how it should feel remains crucial. He mentions developing native Mac apps and tools, like "Trimi," suggesting that programmers still contribute greatly to user experience even within the AI context.
He also speaks of the responsibility developers have in ensuring the safe and effective use of AI tools, by understanding and configuring AI like OpenClaw to prevent bad outcomes. This reflects a crucial role in using agentic AI innovatively and safely.
Steinberger points out the importance of the ecosystem in which developers work, showing that choice and understanding of development environments continue to matter—for example, using Python for model inference or Swift and SwiftUI for deep system integration on Mac, despite challenges.
Furthermore, Steinberger suggests developers' unique skill in understanding how to build will still be in high demand. AI allows people to achieve more, faster due to its continuous improvement, but developers must adapt and continue to grow their expertise to thrive.
Developers Must Adapt Mindset and Skills to Thrive in AI Landscape, Focusing on Design and Problem-Solving
Steinberger and others in the conversation assert that developers must adapt their focus to encompass more complex problem-solving and design-oriented roles as AI systems advance. They suggest p ...
Here’s what you’ll find in our full summary
Future of Programming and Human Developers in AI World
Additional Materials
Clarifications
- Agentic AI refers to artificial intelligence systems that can act autonomously to achieve goals, making decisions and taking actions without constant human input. Unlike traditional AI, which typically follows fixed rules or responds passively, agentic AI can modify its own behavior and software to improve performance. This autonomy allows it to handle complex tasks and adapt dynamically in changing environments. Agentic AI thus functions more like an independent agent than a simple tool.
- When AI "modifies its own software," it means the AI can change its programming or algorithms without human intervention. This allows the AI to improve, adapt, or fix itself autonomously. The implication is a shift from humans writing every line of code to guiding AI systems that evolve independently. It raises new challenges in control, safety, and understanding AI behavior.
- The analogy compares programming to knitting to highlight a shift from necessity to enjoyment. Knitting is often done as a creative hobby rather than a required task. Similarly, programming may become less about routine coding and more about creative expression. This suggests programmers will focus on innovation and artistry rather than repetitive work.
- Traditional programming involves writing explicit code to instruct computers step-by-step. New roles focus on guiding AI systems to generate or modify code, emphasizing design, problem-solving, and decision-making. Developers now act more as supervisors and creative directors rather than manual coders. This shift requires skills in AI interaction, system architecture, and user experience design.
- Directing AI involves defining goals, setting constraints, and guiding AI behavior to achieve desired outcomes. It requires designing workflows, providing feedback, and making high-level decisions rather than writing detailed code. This role focuses on collaboration with AI tools to solve problems creatively and ensure alignment with user needs. Coding, by contrast, is the manual creation of software instructions and algorithms.
- Model inference is the process where a trained AI model makes predictions or decisions based on new data. Python is commonly used for inference because it has extensive libraries and frameworks that simplify running AI models efficiently. Its flexibility and strong community support make it ideal for integrating AI into various applications. This allows developers to quickly test and deploy AI capabilities without building models from scratch.
- Swift is Apple's modern programming language designed for building fast, safe, and expressive applications on Apple platforms. SwiftUI is a framework that allows developers to create user interfaces declaratively, making UI development more intuitive and efficient. Together, they enable deep system integration by providing access to native macOS features and APIs, allowing apps to perform complex tasks and interact closely with the operating system. This combination helps developers create seamless, high-performance Mac applications that feel native and responsive.
- The phrase "human insight, or the 'style, love, that human touch'" refers to the unique creativity, intuition, and emotional understanding humans bring to programming. It means designing software with empathy, aesthetics, and nuanced judgment that AI cannot replicate. This includes anticipating user needs, crafting elegant solutions, and adding personal or cultural context. Practically, it ensures software feels intuitive, meaningful, and aligned with human values.
- Agentic AI refers to AI systems that can make decisions and take actions autonomously, often modifying their own code or behavior. Safety concerns include preventing unintended harmful actions, ensuring AI aligns with human values, and avoiding errors that could cause damage or security risks. Using agentic AI innovatively and safely means designing, monitoring, and controlling these systems to maximize benefits while minimizing risks. Developers must understand AI behavior deeply to configure safeguards and intervene when necessary.
- "Reshaping their identities" means developers must rethink how they see their professional roles beyond just writing code. It involves embracing new skills like guiding AI tools, focusing on design, and solving complex problems. This shift requires adapting to collaborative work with AI rather than solo programming. Ultimately, it calls for a ...
Counterarguments
- AI's ability to modify its own software could lead to unforeseen complexities and errors that require even more sophisticated human intervention.
- The human touch in programming is subjective and may not always lead to better outcomes than AI-generated code.
- Equating programming to knitting may oversimplify the complexities and challenges involved in software development.
- While developers are encouraged to focus on design and problem-solving, not all developers may have the necessary skills or interest in these areas.
- The assertion that developers should not view themselves as specialists might undervalue the importance of deep expertise in certain areas of technology.
- The idea that developers must embrace AI could be seen as disregarding the value of traditional programming skills and the preferences of those who wish to continue coding.
- The responsibility placed on developers to ensure the safe and effective use of AI tools may be unrealistic without proper training, guidelines, and support.
- The claim that developers' unique skills in understanding how to build software will remain in high demand may not account for the possibility of AI systems eventually learning and replicating these skills.
- The suggestion that AI enables people to achieve more and faster overlooks the potential for AI to also introduce new challenges and complexities that could slow down development.
- The emphasis on developers adapting their mindset might not acknowledge the potential resistance or difficulty some individuals may face in transitioning to new roles. ...
Get access to the context and additional materials
Create Summaries for anything on the web
Download the Shortform Chrome extension for your browser
