What Is Your NemoClaw Strategy?
At the start of the AI supercycle that kicked off with ChatGPT, board rooms were terrorized by a single question: “What is your enterprise AI strategy?”
Enterprises move slow. By the time they could formulate an answer, gain some confidence on what is working and what is not, the question shifted to “What is your enterprise agentic strategy?” By the time teams could scramble to piece that together, the question shifted again.
At GTC 2026, Jensen Huang told every CEO they should have an answer to “What is your OpenClaw strategy?” He compared it to Linux. To Kubernetes. To HTTP. The protocols that built the internet.
OpenClaw is the fastest-growing open-source project in GitHub history. 280,000+ stars. More popular than Linux or React. And when Mark Zuckerberg and other tech CEOs saw it, they banned it from their companies. That was the strategy.
Jensen took a different approach. NVIDIA wrapped OpenClaw in enterprise-grade controls and launched NemoClaw at GTC in front of 30,000 attendees. Salesforce, SAP, ServiceNow among the launch partners. Apache 2.0 license. Free. The question sitting in the room was simple: what is your NemoClaw strategy?
Most boards haven’t heard the word yet.
Two years ago the board conversation was simpler. Gen AI is not traditional AI. The tech stack jumps 10x every six months. Workflow compression is the real promise. Agentic commerce and autocoding are the trends to watch. Must start now but run like a marathoner.
Enterprise AI adoption has been a lacklustre story. MIT’s NANDA initiative put a number on it. Ninety-five percent of enterprise AI pilots deliver zero measurable P&L impact.
The earlier question was easier to deal with. You had an enterprise, its objective function, and AI was a tool. Impressive in demos. Summarize a 200-page contract in 30 seconds, everyone claps. Put it in production and it hallucinates the payment terms. Enterprises don’t care about innovation. They care about not looking stupid in front of regulators. In the short term AI increases that risk. So adoption has been fast and slow. Mostly slow.
Now it is not an innovation question. It is a disruption question.
Enterprises are spread across a spectrum on AI.
AI-Augmented: I use AI tools to amplify individual productivity. A developer with Copilot. A marketer with ChatGPT. The tool makes the person faster but the workflow, the org chart, the headcount stays the same. Most enterprises are here. Most think they are further along.
AI-First: I have redefined my workflows around AI, not just added AI to existing workflows. The engineering team ships with three what used to take twelve. The marketing team runs 40 campaigns a quarter instead of 8, and most of them are generated, tested, and killed without a human touching them. Workflows redesigned, not just accelerated. Almost nobody is actually here. Many claim to be. The tell: if your headcount hasn’t changed, you’re not AI-First. You’re AI-Augmented with better slide decks.
AI-Native: I look at my enterprise and say the future of work is human plus claw. In that future, how many agents and how many humans do I need? Nobody is here yet. But this is where the question points.
The AI-Native conversation is a political landmine. Job loss. Workforce restructuring. Every board that touches it wishes they hadn’t. Every board that avoids it is deferring a conversation that gets harder by the quarter.
The first task CEOs face is parsing their teams into three buckets. Savvy folks who will lead with AI. Find them and follow them. People who are open but not self-starters, who need hands-on training. Get them a real program, not certificate business. Then there is the bottom rung. The AI drag. People who will find fault with AI, agents, and claws, say they don’t matter, and refuse to reorganize.
CEOs are being pressured to redesign the workforce even if the AI drag segment is 60% of their organization. CEOs who resist are punished by the market. Doesn’t matter if you are CEO of Adobe or CEO of Walmart. The market doesn’t care about your transition plan. The market cares about your cost structure.
Brutal. Also not new. This is what happened with offshoring. This is what happened with cloud. The playbook is identical. The speed is different.
To understand NemoClaw you have to understand why OpenClaw clicked. In 2025 there were hundreds of candidates trying to be the agentic platform. LangChain, AutoGen, CrewAI, MetaGPT. OpenClaw won.
The journalists will tell you it broke through because of the WhatsApp interface. Or the solo developer story. Or the lobster mascot. All true. None sufficient.
OpenClaw succeeded because of plain-text files in a workspace folder. Not a neural architecture breakthrough. Not a new model. Markdown files. Editable with any text editor. Version-controllable with Git. No database. No config panel.
Eight files boot every session. Four carry the weight.
SOUL.md defines who the agent is. Personality, values, behavioral boundaries. The first thing the agent reads every time it wakes. A typical SOUL.md opens with “You’re not a chatbot. You’re becoming someone.” Not what to do. Who to be. Without it, the agent is a raw language model with no persistent identity. With it, the agent has a character that survives across sessions, across model swaps, across reboots.
MEMORY.md is the agent’s long-term memory. Curated preferences, hard-won rules, patterns distilled from weeks of daily logs. The agent writes daily notes to memory/YYYY-MM-DD.md files, and the important patterns get promoted into MEMORY.md over time. Six months of accumulated context that no fresh installation can replicate.
HEARTBEAT.md is the scheduling brain. Every 30 minutes, the agent wakes up, reads its schedule, and acts. Check Google Calendar. Send a Telegram summary. Verify SSL certs. Monitor disk usage. Run a weekly ticket report every Monday at 8am. Nobody prompted it. Nobody asked. The flip from human-initiated to beat-initiated. No other framework made that leap.
AGENTS.md is the operating manual. If SOUL.md answers “who are you,” AGENTS.md answers “what do you do and how.” Workflow instructions, procedural rules, memory management protocols, safety constraints. The rules the agent follows even when it is tempted to over-explain or cut corners.
The remaining four handle configuration: IDENTITY.md (name, role, avatar), USER.md (who the human operator is, their preferences, their timezone), TOOLS.md (what tools are available, SSH hosts, device IDs), and BOOTSTRAP.md (first-run onboarding script, deleted after use).
Other frameworks treated the agent as a tool you invoke. OpenClaw treated it as an entity that persists.
Startup CEO I work with, Rushant Ashtputre, put it well: a Claw is like a Pokemon. I had called it a Transformer Autobot that can shape-shift, which gives it both its power and its danger. The Pokemon metaphor captures something the technical descriptions miss. The personality. The emotional alignment. The proactiveness. That is a step jump from everything that came before.
The model (Claude, GPT, Kimi K2) is the brain. The markdown files are the personality. And underneath both sits Pi, a minimal coding agent runtime written by Mario Zechner. Pi handles the agent loop: session management, tool execution, model routing, context compaction. Pi is the harness.
OpenClaw is the personality layer built on top of the harness. Swap the brain. The agent persists. The agent isn’t the model. The agent is the harness.
The brain is commodity. GPT-4 API prices dropped 95% in 18 months, from $30 per million tokens to $1.50. The harness accumulates. MEMORY.md grows every day. SOUL.md gets refined. Skills compound. Heartbeat tasks multiply. Six months in, the agent knows which Slack channels its operator ignores, which meeting invites are optional, which vendor emails need a reply within the hour and which can wait until Thursday. No fresh installation replicates that. No competitor can purchase it.
When the processing layer commoditizes faster than the context layer accumulates, value migrates to context. Models commoditize every six months. Context accumulates every day. Very few teams are optimizing for this yet.
If OpenClaw is a Pokemon, NemoClaw is the corporate-sanctioned version. OpenShell sandbox with Landlock kernel isolation, YAML-defined governance policies, session persistence across reboots, agent identity that maps to your org chart. The brain is whatever model you choose: Nemotron running locally on a DGX Spark for sensitive data, Claude or GPT via API for everything else. The brain is the part you can swap on Tuesday.
NemoClaw strategy is really a parallel organization strategy. Humans plus AI entities. Entities with identity, memory, autonomy, and rules. Most boards are still debating which LLM vendor to standardize on. That’s not the constraint.
There’s a version of this where none of it works the way we expect. Most enterprise AI bets have gone that way so far.
If you sit in a board meeting today, the same questions keep coming up.
The org question shows up quickly. What does a parallel org even look like? A NemoClaw FTE-equivalent handles repeatable workflows at machine speed with full audit trails. Procurement processing: matching POs to invoices, flagging discrepancies, routing approvals. Vendor onboarding: collecting W-9s, verifying insurance certificates, populating ERP records. Compliance monitoring: scanning contract amendments for regulatory triggers. Tier-1 support triage: classifying tickets, pulling knowledge base answers, escalating edge cases to humans. Workflows where the process is documented, the exceptions are bounded, and the cost of a human doing it is $150-250K per year fully loaded.
Then someone asks about roles. Which ones go first. Which five roles in your organization could be NemoClaw FTE-equivalents within 12 months? The CFO already thought about this. The board hasn’t.
And eventually it lands on timing. What happens if we wait?
In 2023, waiting didn’t cost much. 2024, still manageable. Now it compounds.
Every month a NemoClaw agent operates inside an enterprise, it accumulates institutional context that no new deployment can replicate. The undocumented exception in the procurement workflow where orders above $50K from APAC vendors need a second sign-off that isn’t in the policy manual. The compliance quirk that legal added in 2023 and never told engineering about. The VP who needs three touches before she approves and will reject on the second if the tone is wrong.
Every day an agent operates, it learns things about your organization that no fresh deployment can replicate. That accumulated context is the advantage, not which AI model you picked. If competitors deploy NemoClaw FTE-equivalents at scale in 2026, catching up in 2027 requires buying something that isn’t for sale.
And the conversation nobody wants to have. One NemoClaw agent costs less per month than one employee costs per week.
Every technology transition has its euphemism. Cloud was “digital transformation” while 40% of on-prem sysadmin roles disappeared between 2012 and 2018. Offshoring was “global delivery” while Bangalore absorbed 3.7 million IT services jobs that used to sit in New Jersey and Ohio. AI agents will be “the parallel organization” and “augmentation.”
Everyone calls it augmentation. It’s displacement. Just delayed. Technology that makes a human 10x more productive does not create 10x more work for that human. It creates the same work done by fewer humans. The delay is the political cover. We have done this before. We will do it again. What changes is how honest the organization is willing to be about it.
Board members have been through this. Cloud displaced on-prem IT teams. SaaS displaced custom software teams. Every time, the cheerleaders said “it’s about augmenting, not replacing.” Every time, headcount went down 18 months later. If this piece sounds like the cheerleaders, you should stop reading it.
The parallel organization means fewer humans doing more, with agents handling the repeatable work. It’s going to happen anyway. The only difference is whether it’s planned.
Intention looks like this: the procurement specialist who spent 70% of her time matching POs to invoices now spends that time negotiating payment terms, resolving vendor disputes, managing relationships with suppliers who have leverage. Exception handling. Judgment calls. The gap between documented process and actual process.
Accident looks like this: you deploy agents, discover the humans are redundant, and do layoffs with no transition plan. This is how every previous technology transition went wrong. Not a mystery. Just a pattern nobody wants to recognize while they’re inside it.
The redeployment test is simple. Before approving any NemoClaw deployment, answer one question: where does the human currently doing this work go next? If the answer is “we’ll figure that out later,” that is not a NemoClaw strategy. That is a headcount reduction dressed in AI language. It is also the most likely outcome, because it is always the most likely outcome.
Deploy first. Prove the agent works at production quality for six months. Then redesign the org. The board that deploys agents and lays off people in the same quarter is making an irreversible bet on a reversible experiment. Most boards will do exactly this. The ones that don’t will be worth studying.
NemoClaw is alpha software. Not production-ready.
Four questions in three years. Each arrived faster than the last. Each harder. This used to be about tools. Then workflows. Now it’s about people. That’s where it gets uncomfortable.
Tools are easy. Procurement handles it. Workflows take longer. Ops gets involved. People is where things slow down and where the real resistance lives. That’s where this actually gets hard.
PS: More detailed thoughts in my book with @Kashi KS https://www.amazon.com/What-your-NemoClaw-Strategy-Competition-ebook/dp/B0GGH366DB/