PJFP.com

Pursuit of Joy, Fulfillment, and Purpose

Tag: coding automation

  • OpenClaw & The Age of the Lobster: How Peter Steinberger Broken the Internet with Agentic AI

    In the history of open-source software, few projects have exploded with the velocity, chaos, and sheer “weirdness” of OpenClaw. What began as a one-hour prototype by a developer frustrated with existing AI tools has morphed into the fastest-growing repository in GitHub history, amassing over 180,000 stars in a matter of months.

    But OpenClaw isn’t just a tool; it is a cultural moment. It’s a story about “Space Lobsters,” trademark wars with billion-dollar labs, the death of traditional apps, and a fundamental shift in what it means to be a programmer. In a marathon conversation on the Lex Fridman Podcast, creator Peter Steinberger pulled back the curtain on the “Age of the Lobster.”

    Here is the definitive deep dive into the viral AI agent that is rewriting the rules of software.

    The TL;DW (Too Long; Didn’t Watch)

    • The “Magic” Moment: OpenClaw started as a simple WhatsApp-to-CLI bridge. It went viral when the agent—without being coded to do so—figured out how to process an audio file by inspecting headers, converting it with ffmpeg, and transcribing it via API, all autonomously.
    • Agentic Engineering > Vibe Coding: Steinberger rejects the term “vibe coding” as a slur. He practices “Agentic Engineering”—a method of empathizing with the AI, treating it like a junior developer who lacks context but has infinite potential.
    • The “Molt” Wars: The project survived a brutal trademark dispute with Anthropic (creators of Claude). During a forced rename to “MoltBot,” crypto scammers sniped Steinberger’s domains and usernames in seconds, serving malware to users. This led to a “Manhattan Project” style secret operation to rebrand as OpenClaw.
    • The End of the App Economy: Steinberger predicts 80% of apps will disappear. Why use a calendar app or a food delivery GUI when your agent can just “do it” via API or browser automation? Apps will devolve into “slow APIs”.
    • Self-Modifying Code: OpenClaw can rewrite its own source code to fix bugs or add features, a concept Steinberger calls “self-introspection.”

    The Origin: Prompting a Revolution into Existence

    The story of OpenClaw is one of frustration. In late 2025, Steinberger wanted a personal assistant that could actually do things—not just chat, but interact with his files, his calendar, and his life. When he realized the big AI labs weren’t building it fast enough, he decided to “prompt it into existence”.

    The One-Hour Prototype

    The first version was built in a single hour. It was a “thin line” connecting WhatsApp to a Command Line Interface (CLI) running on his machine.

    “I sent it a message, and a typing indicator appeared. I didn’t build that… I literally went, ‘How the f*** did he do that?’”

    The agent had received an audio file (an opus file with no extension). Instead of crashing, it analyzed the file header, realized it needed `ffmpeg`, found it wasn’t installed, used `curl` to send it to OpenAI’s Whisper API, and replied to Peter. It did all this autonomously. That was the spark that proved this wasn’t just a chatbot—it was an agent with problem-solving capabilities.

    The Philosophy of the Lobster: Why OpenClaw Won

    In a sea of corporate, sanitized AI tools, OpenClaw won because it was weird.

    Peter intentionally infused the project with “soul.” While tools like GitHub Copilot or ChatGPT are designed to be helpful but sterile, OpenClaw (originally “Claude’s,” a play on “Claws”) was designed to be a “Space Lobster in a TARDIS”.

    The soul.md File

    At the heart of OpenClaw’s personality is a file called soul.md. This is the agent’s constitution. Unlike Anthropic’s “Constitutional AI,” which is hidden, OpenClaw’s soul is modifiable. It even wrote its own existential disclaimer:

    “I don’t remember previous sessions… If you’re reading this in a future session, hello. I wrote this, but I won’t remember writing it. It’s okay. The words are still mine.”

    This mix of high-utility code and “high-art slop” created a cult following. It wasn’t just software; it was a character.

    The “Molt” Saga: A Trademark War & Crypto Snipers

    The projects massive success drew the attention of Anthropic, the creators of the “Claude” model. They politely requested a name change to avoid confusion. What should have been a simple rebrand turned into a cybersecurity nightmare.

    The 5-Second Snipe

    Peter attempted to rename the project to “MoltBot.” He had two browser windows open to execute the switch. In the five seconds it took to move his mouse from one window to another, crypto scammers “sniped” the account name.

    Suddenly, the official repo was serving malware and promoting scam tokens. “Everything that could go wrong, did go wrong,” Steinberger recalled. The scammers even sniped the NPM package in the minute it took to upload the new version.

    The Manhattan Project

    To fix this, Peter had to go dark. He planned the rename to “OpenClaw” like a military operation. He set up a “war room,” created decoy names to throw off the snipers, and coordinated with contacts at GitHub and X (Twitter) to ensure the switch was atomic. He even called Sam Altman personally to check if “OpenClaw” would cause issues with OpenAI (it didn’t).

    Agentic Engineering vs. “Vibe Coding”

    Steinberger offers a crucial distinction for developers entering this new era. He rejects the term “vibe coding” (coding by feel without understanding) and proposes Agentic Engineering.

    The Empathy Gap

    Successful Agentic Engineering requires empathy for the model.

    • Tabula Rasa: The agent starts every session with zero context. It doesn’t know your architecture or your variable names.
    • The Junior Dev Analogy: You must guide it like a talented junior developer. Point it to the right files. Don’t expect it to know the whole codebase instantly.
    • Self-Correction: Peter often asks the agent, “Now that you built it, what would you refactor?” The agent, having “felt” the pain of the build, often identifies optimizations it couldn’t see at the start.

    Codex (German) vs. Opus (American)

    Peter dropped a hilarious but accurate analogy for the two leading models:

    • Claude Opus 4.6: The “American” colleague. Charismatic, eager to please, says “You’re absolutely right!” too often, and is great for roleplay and creative tasks.
    • GPT-5.3 Codex: The “German” engineer. Dry, sits in the corner, doesn’t talk much, reads a lot of documentation, but gets the job done reliably without the fluff.

    The End of Apps & The Future of Software

    Perhaps the most disruptive insight from the interview is Steinberger’s view on the app economy.

    “Why do I need a UI?”

    He argues that 80% of apps will disappear. If an agent has access to your location, your health data, and your preferences, why do you need to open MyFitnessPal? The agent can just log your calories based on where you ate. Why open Uber Eats? Just tell the agent “Get me lunch.”

    Apps that try to block agents (like X/Twitter clipping API access) are fighting a losing battle. “If I can access it in the browser, it’s an API. It’s just a slow API,” Peter notes. OpenClaw uses tools like Playwright to simply click “I am not a robot” buttons and scrape the data it needs, regardless of developer intent.

    Thoughts: The “Mourning” of the Craft

    Steinberger touched on a poignant topic for developers: the grief of losing the craft of coding. For decades, programmers have derived identity from their ability to write syntax. As AI takes over the implementation, that identity is under threat.

    But Peter frames this not as an end, but an evolution. We are moving from “programmers” to “builders.” The barrier to entry has collapsed. The bottleneck is no longer your ability to write Rust or C++; it is your ability to imagine a system and guide an agent to build it. We are entering the age of the System Architect, where one person can do the work of a ten-person team.

    OpenClaw is not just a tool; it is the first true operating system for this new reality.

  • Composer: Building a Fast Frontier Model with Reinforcement Learning

    Composer represents Cursor’s most ambitious step yet toward a new generation of intelligent, high-speed coding agents. Built through deep reinforcement learning (RL) and large-scale infrastructure, Composer delivers frontier-level results at speeds up to four times faster than comparable models:contentReference[oaicite:0]{index=0}. It isn’t just another large language model; it’s an actively trained software engineering assistant optimized to think, plan, and code with precision — in real time.

    From Cheetah to Composer: The Evolution of Speed

    The origins of Composer go back to an experimental prototype called Cheetah, an agent Cursor developed to study how much faster coding models could get before hitting usability limits. Developers consistently preferred the speed and fluidity of an agent that responded instantly, keeping them “in flow.” Cheetah proved the concept, but it was Composer that matured it — integrating reinforcement learning and mixture-of-experts (MoE) architecture to achieve both speed and intelligence.

    Composer’s training goal was simple but demanding: make the model capable of solving real-world programming challenges in real codebases using actual developer tools. During RL, Composer was given tasks like editing files, running terminal commands, performing semantic searches, or refactoring code. Its objective wasn’t just to get the right answer — it was to work efficiently, using minimal steps, adhering to existing abstractions, and maintaining code quality:contentReference[oaicite:1]{index=1}.

    Training on Real Engineering Environments

    Rather than relying on synthetic datasets or static benchmarks, Cursor trained Composer within a dynamic software environment. Every RL episode simulated an authentic engineering workflow — debugging, writing unit tests, applying linter fixes, and performing large-scale refactors. Over time, Composer developed behaviors that mirror an experienced developer’s workflow. It learned when to open a file, when to search globally, and when to execute a command rather than speculate.

    Cursor’s evaluation framework, Cursor Bench, measures progress by realism rather than abstract metrics. It compiles actual agent requests from engineers and compares Composer’s solutions to human-curated optimal responses. This lets Cursor measure not just correctness, but also how well the model respects a team’s architecture, naming conventions, and software practices — metrics that matter in production environments.

    Reinforcement Learning as a Performance Engine

    Reinforcement learning is at the heart of Composer’s performance. Unlike supervised fine-tuning, which simply mimics examples, RL rewards Composer for producing high-quality, efficient, and contextually relevant work. It actively learns to choose the right tools, minimize unnecessary output, and exploit parallelism across tasks. The model was even rewarded for avoiding unsupported claims — pushing it to generate more verifiable and responsible code suggestions.

    As RL progressed, emergent behaviors appeared. Composer began autonomously running semantic searches to explore codebases, fixing linter errors, and even generating and executing tests to validate its own work. These self-taught habits transformed it from a passive text generator into an active agent capable of iterative reasoning.

    Infrastructure at Scale: Thousands of Sandboxed Agents

    Behind Composer’s intelligence is a massive engineering effort. Training large MoE models efficiently requires significant parallelization and precision management. Cursor’s infrastructure, built with PyTorch and Ray, powers asynchronous RL at scale. Their system supports thousands of simultaneous environments, each a sandboxed virtual workspace where Composer experiments safely with file edits, code execution, and search queries.

    To achieve this scale, the team integrated MXFP8 MoE kernels with expert and hybrid-sharded data parallelism. This setup allows distributed training across thousands of NVIDIA GPUs with minimal communication cost — effectively combining speed, scale, and precision. MXFP8 also enables faster inference without any need for post-training quantization, giving developers real-world performance gains instantly.

    Cursor’s infrastructure can spawn hundreds of thousands of concurrent sandboxed coding environments. This capability, adapted from their Background Agents system, was essential to unify RL experiments with production-grade conditions. It ensures that Composer’s training environment matches the complexity of real-world coding, creating a model genuinely optimized for developer workflows.

    The Cursor Bench and What “Frontier” Means

    Composer’s benchmark performance earned it a place in what Cursor calls the “Fast Frontier” class — models designed for efficient inference while maintaining top-tier quality. This group includes systems like Haiku 4.5 and Gemini Flash 2.5. While GPT-5 and Sonnet 4.5 remain the strongest overall, Composer outperforms nearly every open-weight model, including Qwen Coder and GLM 4.6:contentReference[oaicite:2]{index=2}. In tokens-per-second performance, Composer’s throughput is among the highest ever measured under the standardized Anthropic tokenizer.

    Built by Developers, for Developers

    Composer isn’t just research — it’s in daily use inside Cursor. Engineers rely on it for their own development, using it to edit code, manage large repositories, and explore unfamiliar projects. This internal dogfooding loop means Composer is constantly tested and improved in real production contexts. Its success is measured by one thing: whether it helps developers get more done, faster, and with fewer interruptions.

    Cursor’s goal isn’t to replace developers, but to enhance them — providing an assistant that acts as an extension of their workflow. By combining fast inference, contextual understanding, and reinforcement learning, Composer turns AI from a static completion tool into a real collaborator.

    Wrap Up

    Composer represents a milestone in AI-assisted software engineering. It demonstrates that reinforcement learning, when applied at scale with the right infrastructure and metrics, can produce agents that are not only faster but also more disciplined, efficient, and trustworthy. For developers, it’s a step toward a future where coding feels as seamless and interactive as conversation — powered by an agent that truly understands how to build software.