From Flo's AI Lab
Two days inside an AI-first innovation consultancy in Kassel this week. This is a firm with 25 years of experience, 68 design awards, and deep expertise in HMI and industrial design. The brief: showcase the possibilities of claude coworker and skills. Surprise on day one: the team assumed Claude Coworker runs inside a browser tab. When they saw it physically operating their Mac, opening apps, they understood the possibilities. We built three custom skills for the team (meeting assist, market research, and content creation). Day two, we built a complete content agent: a LinkedIn posting skill trained on 1,400 of the founder's posts to match tone of voice, a newsletter pipeline with automated review, and Gemini-powered image generation in their house style. What started as vibe coding experiments became structured agent workflows that the team can run themselves.
What happened in that Kassel workshop is a microcosm of something much larger. This week, the entire AI industry converged on a single architectural pattern. And if you understand the pattern, you understand where every software product is heading.
The Architecture That Will Eat Software
Peter Steinberger may have created the dominant design for AI agents. His open-source project OpenClaw combines an LLM for reasoning, pluggable skills for real-world actions, a scheduler that acts without being prompted, and persistent local memory. That four-part pattern has exploded to over 300,000 GitHub stars, making it one of the fastest-growing open-source projects in history (Fortune, Feb 2026). In February, Steinberger joined OpenAI to build "next generation personal agents," while OpenClaw itself moves to an independent foundation (TechCrunch, Feb 2026).
At NVIDIA's GTC 2026 keynote last week, Jensen Huang called OpenClaw "the most popular open-source project in the history of humanity" and declared that "every company needs an OpenClaw strategy." He compared it to Linux, to Kubernetes, to HTML: infrastructure that arrives at the right moment for an entire industry to build on (Fierce Network, Mar 2026, 36kr, Mar 2026).
Why is Huang giving a solo developer's side project the Linux treatment? Because OpenClaw codifies a four-part architecture that every major AI company is converging on independently:
A brain (any LLM: Claude, GPT, DeepSeek, or a local model). Skills and tools (thousands of plugins that let the agent interact with the real world). A scheduler that wakes the agent on its own, checks the task list, and acts without prompting. Persistent memory that accumulates context over time and stays on your own hardware.
Brain, skills, scheduler, memory. Four components. One pattern. And everyone is building it:
Anthropic rolled out persistent memory for all Claude users, launched Claude Code Channels for Telegram and Discord (VentureBeat, Mar 2026), and introduced scheduled tasks that run autonomously in the cloud. Perplexity launched "Computer," a multi-model agent coordinating 19 LLMs that creates sub-agents and runs continuously on local hardware (TechCrunch, Feb 2026). NVIDIA announced NemoClaw, an enterprise stack that wraps OpenClaw with security controls and its own Nemotron models (NVIDIA Newsroom, Mar 2026). The chip company that powers the infrastructure is now shipping the agent runtime.
This is convergence, not coincidence. In 1945, John von Neumann's "First Draft of a Report on the EDVAC" defined how a computer should work: a processing unit, memory, and a control unit. He did not invent any component. He defined how they fit together, and every computer since follows that blueprint. The same thing is happening now. Every AI company is arriving at the same four-part agent architecture because it is the right decomposition of autonomous software.
What This Means: Software Becomes Staff
If this architecture becomes as universal as von Neumann's, the implication is not "better software." The implication is that software stops being a tool and starts being a colleague.
I reached this exact conclusion six months ago while building Taxfinito, an AI-powered tax and accounting system. The original brief was "build better accounting software." What I actually built was a virtual accountant: an agent with an LLM brain, skills for document processing and compliance checks, a scheduler that proactively handles deadlines, and memory that accumulates client context over time. The software does not wait for a user to click buttons. It processes incoming invoices, checks GoBD compliance, flags anomalies, and prepares filings on its own schedule.
That is not a feature upgrade. That is a category shift. Agentic engineering turns software from a passive tool into an active participant. And it applies everywhere:
Not HR software. A Head of HR who screens candidates, schedules interviews, tracks compliance deadlines, and drafts offer letters proactively.
Not a legal text chat. A Legal Counsel who monitors regulatory changes, reviews contracts against precedent, flags risk clauses, and prepares board briefings on a schedule.
Not tax preparation help. A tax adviser who knows your situation, watches for relevant rulings, optimises your position continuously, and files when the time is right.
Clayton Christensen would call this a jobs-to-be-done disruption (The Innovator's Solution, 2003). The job was never "use accounting software." The job was "have someone competent handle my books." The software was a compromise because the real solution (a dedicated professional) was too expensive for most businesses. When an agent can do the job at software prices, the compromise dissolves. Every SaaS category is a job that was too expensive to hire for. AI agents remove that constraint.
What This Means Monday Morning
For CEOs: Audit your software stack not by features, but by jobs. For each tool, ask: "Is this a tool I operate, or could this be a virtual team member that operates itself?" The products that adopt the brain-skills-scheduler-memory pattern will deliver fundamentally more value than those that remain passive dashboards waiting for clicks. Prioritise vendors who are building agents rather than just adding AI features to existing interfaces.
For PE managers: The "AI-native" label is about to get a precise definition. Any portfolio company whose product follows the four-component pattern (LLM reasoning, extensible skills, proactive scheduling, persistent memory) is building in the right architectural direction. Companies still shipping traditional SaaS with a chatbot bolted on are one OpenClaw plugin away from disruption. Steinberger building this architecture inside OpenAI, with NVIDIA shipping enterprise tooling around it, means distribution is a matter of months, not years.
For service providers: The Kassel workshop taught me something. Design agencies, law firms, and consultancies: your value lies not in production work. It is in the judgment, the client relationship, and the strategic framing. Build your own agents for the production layer. Use OpenClaw, claude code, or whatever fits your stack. The firms that treat this as a core competency will offer services that firms without it simply cannot match. Vibe coding got you started. The agent architecture gets you to scale.
What I'm Watching
Jensen Huang's comparison of OpenClaw to Linux is either visionary or premature. Linux took 30 years to reach ubiquity. OpenClaw hit 300,000 stars in weeks. But stars are not adoption, and adoption is not durability. Early security vulnerabilities exposed over 135,000 systems to the open internet (TechCrunch, Mar 2026). That gap between viral growth and enterprise readiness is exactly what NemoClaw is designed to close. If NVIDIA's bet pays off, the agent layer becomes infrastructure: invisible, essential, and open.
If this changed how you think about your software stack or your next product roadmap, forward it to someone still calling chatbots "AI strategy."
Want to explore how AI agents can reposition your business? Visit /angebote.