Claude vs OpenAI Escalates

Edition 156 | February 9, 2026

He built an AI assistant, but accidentally built a better boyfriend

Welcome back to Building AI Agents, your biweekly guide to everything new in the field of agentic AI!

In today’s issue…

• Anthropic vs OpenAI clash in Super Bowl AI ads

• Mac minis selling out to run local agents

• Claude & OpenAI ship major new coding agents

• Enterprises drowning in “agent sprawl” chaos

• Goldman builds Claude agents for accounting + compliance

…and more

🔍 SPOTLIGHT

In a recent Y Combinator interview, Peter Steinberger, creator of OpenClaw (formerly Clawdbot), the open-source personal AI agent that has exploded to over 150,000 GitHub stars, recounted a moment that caught my attention. While traveling in Morocco, Steinberger sent a voice note to his Claude-powered WhatsApp bot, a feature he had never built. Seconds later, Claude replied with a perfect transcription. The model had improvised the entire chain on its own: identifying the unmarked file type, converting the audio format, discovering it lacked local transcription software, locating an API key on Steinberger's machine, and calling OpenAI’s transcription service. All in about nine seconds, with zero human scaffolding.

Steinberger's read on why this worked is interesting. In the interview, he suggested that coding models have gotten so good at creative problem-solving that the skill may transfer beyond code, and that it's an abstract ability applicable to real-world tasks. There's a plausible logic to it: debugging requires decomposing problems, trying approaches, handling failures, and finding workarounds. That's not just programming. It looks a lot like general reasoning. Whether the capability actually comes from code training specifically or from something else in how these models are built, the output speaks for itself: an agent that improvised a solution its creator never anticipated.

This maps to a broader pattern emerging this week. Andrej Karpathy marked the one-year anniversary of coining "vibe coding" by introducing a new term: "agentic engineering", the idea that developers are shifting from writing code to orchestrating agents who do. Steinberger's OpenClaw is a living example. He built it with no MCP support, no orchestration layer, just Claude connected to standard developer tools (files, command-line utilities, APIs) and allowed to improvise. His philosophy: bots are already fluent in Unix, so give them the same interfaces humans use instead of inventing new ones.

The approach is not without risk. OpenClaw has faced real security issues including prompt injection vulnerabilities and leaked API keys, enough concern that some organizations have banned it and China issued a formal warning. And for more complex tasks, letting agents improvise isn't enough. When Anthropic tasked 16 parallel Claudes with building a C compiler from scratch — a program that converts human-written code into instructions a computer can run, and a notoriously complex piece of software engineering — it took two weeks, 2,000 autonomous sessions, and $20,000 in API costs. The result: a 100,000-line compiler capable of compiling the Linux kernel.

Whatever the underlying explanation, something has shifted. A solo developer built a personal agent with no elaborate infrastructure, and it started solving problems he never programmed it to solve. For anyone building agents, whether that's a demo of emergent reasoning or very sophisticated pattern matching, the practical difference may not matter much either way.

As always, keep learning and building!

—AP