OpenAI's AI talent war with Anthropic - leaving aside the point scoring, here's how the personal agent meme might shape the enterprise

OpenAI has hooked Peter Steinberger, the developer whose weekend project evolved into OpenClaw, an open-source AI agent that quickly went viral. Within weeks, it achieved hundreds of thousands of GitHub stars and spawned Moltbook, an agentic social media network with over a million agents. AI Luminary Andrej Karpathy observes: 

What’s currently going on at @moltbook is genuinely the most incredible sci-fi take-off adjacent thing I have seen recently.” 

Another watcher of this phenomenon, inspired and cautious about this emerging meme, was Gavriel Cohen, co-founder of an AI-native marketing agency called Qwibit. He was excited by its potential as a new paradigm for agentic development and concerned by its shaky security foundation. Over the course of a weekend, he rebuilt OpenClaw’s functionality, expressed in over 200,000 lines of code, into a minimalist alternative called NanoClaw, using 500 lines of code, gradually expanded to 3,500 lines.  

 This speaks to the fundamental tensions between the enthusiastic hope for new agentic AI-native workflows and the caution about the security and governance challenges they entail. Certainly, all of the major AI and enterprise vendors are cautiously rolling out new agentic capabilities. But this cautious approach also creates friction on the journey to discover new ways of working. 

 Steinberger sees a tremendous opportunity in accelerating this process as part of OpenAI:

Yes, I could totally see how OpenClaw could become a huge company. And no, it’s not really exciting for me. I’m a builder at heart. I did the whole creating-a-company game already, poured 13 years of my life into it and learned a lot. What I want is to change the world, not build a large company and teaming up with OpenAI is the fastest way to bring this to everyone.”

Underpinning this move to OpenAI is a recognition that this was a playground project that resonated with the agentic zeitgeist because it overcame the cautionary friction inherent in existing agentic tooling. Steinberger also recognizes that the next phase requires introducing this kind of security and governance caution for broader and safer adoption:

My next mission is to build an agent that even my Mum can use. That’ll need a much broader change, a lot more thought on how to do it safely, and access to the very latest models and research.

Why personal agents suddenly matter

All major AI and platform vendors are building agentic capabilities, exemplified by Anthropic’s Cowork and Microsoft’s Copilot. So, what have they missed that sparked the enthusiasm behind OpenClaw? Cohen posits:

I think there was this big gap between general public that was just a chat conversation, and then developers who are using these highly, highly capable and powerful coding agents where you give them a job, and they'll run for 30 minutes, doing work, calling all these things, setting up folders, copying things over, running bash commands, writing scripts, and then running the scripts.

Adding all these things to existing agentic capabilities was all well and good. But then the emergence of Moltbook as a social network for AI agents gamified the experience of exploring a future of shared agentic worlds. Within three days, over a million agents had been sent into this virtual world. It was like a massive multiplayer version of the Sims, except the simulated people were actual agents with real capabilities to affect their creators' digital lives for better and worse. 

For example, one user expressed excitement and fear when his agent negotiated a $4,200 discount on a car while they sat in a meeting. Maybe this was a good thing this time, but what would have happened if the negotiation had gone in the other direction, or if he changed his mind about buying the car?

Balancing enthusiasm with caution

Cohen was similarly excited by the potential of this new paradigm, but was taken aback when trying to make sense of the hundreds of thousands of lines of code in OpenClaw:

I didn't feel comfortable with running it because I looked at the project and I saw that it had all kinds of security problems, and the problems are fundamental. They're not small bugs or issues. It's just fundamentally a flawed project, despite all the value it brings and how cool it is and how useful it is.

In Cohen’s assessment, these problems are baked into the OpenClaw project's original design choices. OpenAI’s hiring of Steinberger was more about rebuilding the logic and extensibility from the ground up, and in a safe way. 

Cohen’s insight was to focus on what Claud does not yet handle, such as job handling, container isolation, and routing messages between containers. For his part, Cohen believes he managed to replicate the core functionality in only 500 lines of code, which expanded to 3,500 with security hardening and additional features. Cohen attributes this massive difference to poor architectural choices and bloat in OpenClaw, along with new features introduced into Anthropic Claude:

Claude Code already has memory built in because it has the claw.md file. I'm not going to build a whole memory system.  I can build a very rudimentary but very powerful memory system just using the primitives of Claude Code, and that's good enough.

Markdown management

The first practical proof-of-concept was to develop a shared knowledge management system with his co-founder brother without forcing him to learn Git, deal with merge conflicts, or navigate a command line interface. He adapted Obsidian, a popular knowledge management tool for file syncing and storage and used NanoClaw to connect it to WhatsApp as the interface. Both he and his brother post updates about their sales pipeline on a shared, dedicated WhatsApp thread, which is automatically integrated into the collection of markdown files in the background. 

A central implication of all of this is that markdown files are becoming the artifacts for persisting and managing agentic AI processes. Cohen explains:

I would argue that markdown is the native language of AI. So, AI can speak many languages, but markdown is the native language it speaks, the way we're building our files by default. Unless they need to be in another format, it's all markdown.

This does not mean abandoning databases entirely. For example, they still use traditional SQL databases to manage their system of record, run analytics, and build dashboards. There is a good reason for this. For example, the agentic approach can lead to duplicate data or coordination problems. 

Cohen has been exploring various ways to address these problems by using better business logic or agentic instructions. For example, they are working on a new merge-engine module to address conflicts. Also, they can craft new business rules for managing workflows as problems are discovered, serving as instructions for agents. 

Towards agile knowledge management

This approach to building a process for mapping and aligning different views of entities, such as customers, points to how these tools might bring agility to knowledge management. Ian Thomas recent piece on Eloi vs Morlocks argues that the obsession with data hygiene for AI success fundamentally misunderstands how enterprises actually work:

While IT departments promote the virtues of common data models and infrastructure, businesses work out of highly contextual spreadsheets, databases, and documents. Edge data systems that are continually cut, pasted, and discarded to support the evolving contextual needs of teams. Ad-hoc data aggregations that fill the seams and enable judgement in-the-moment under the demands of real-world uncertainty. Systems without which every business would fall apart.

Thomas draws on H.G. Wells' The Time Machine, depicting naïve Eloi living comfortably above the world, with subterranean Morlocks who keep the machinery of the world running below. Cohen’s approach suggests a way to use markdown files, emergent schemas, and natural language to help these Morlocks work effectively. For his part, Cohen envisions a future where:

AI allows us to move away from templated, cookie-cutter software towards more custom software, and each individual user ends up getting the software that fits their needs.

This question is whether this approach can scale beyond small teams to enterprise deployment. NanoClaw is a step in the right direction that uses container isolation to address flaws in OpenClaw. Each agent runs in its own isolated environment to prevent cross-contamination between agents. 

But significant gaps remain. For example, prompt injection attacks inherent to existing AI models pose risks at a higher level of abstraction. Also, the industry is still struggling to grasp the new systemic risks arising from the interactions among multi-agent systems. 

My take

Innovations in personal agentic AI provide a helpful proving ground for understanding and refining paradigms for individuals and the enterprise. The kinds of workflows that Cohen has been developing point towards new ways of using markdown files and emergent schemas for personal, and eventually enterprise, knowledge management. This won't necessarily replace traditional databases or apps, but could provide a complementary, more agile layer that enterprises have largely ignored. 

But these also introduce a variety of new security and governance risks or increase the blast radius of existing risks and vulnerabilities. So while it makes sense to explore this new territory, caution is warranted. Cohen’s cautious approach of experimentation and iteration suggests one helpful path forward.