Franziska Hinkelmann, PhD

Your agent is only as trustworthy as the environment it runs in. So today we launch something new with @NVIDIA. AI agents have gone from prompt-and-response tools to autonomous systems that run for hours, write their own code, build their own tools, and learn as they go. The OpenClaw project earlier this year made this concrete, self-evolving agents that plan complex tasks, generate their own tools, and run continuous workflows. We built CrewAI for exactly this. Long-running multi-agent systems. Persistent memory. A dual-layer architecture where Flows handle deterministic control, and Crews handle reasoning. Developers get precise control over how much autonomy each part of the system gets. But here's what keeps coming up with enterprise teams. When an agent can install packages, write files, and generate its own tools, it can also do things you didn't plan for. Most agents inherit the full permissions of whoever launched them. Security checks are usually built inside the agent — so a self-evolving agent could, in theory, work around its own guardrails. This is the trust gap. The real reason most enterprise agent projects don't make it to production. CrewAI addresses a lot of this at the orchestration layer: guardrails, human-in-the-loop, and hierarchical task scoping. But orchestration alone can't close the full gap. You also need enforcement at the infrastructure level, below the agent, where the agent can't reach. That's why we're working with NVIDIA on NemoClaw. NVIDIA NemoClaw is an open-source stack that simplifies running OpenClaw always-on assistants safely, with a single command. It includes the NVIDIA OpenShell Runtime with three core capabilities: A sandbox for isolated execution — agents operate freely without affecting the host. A policy engine that evaluates every action at the binary, destination, and network level. A privacy router that directs inference to local or external models based on your enterprise policies. The critical design choice: enforcement happens at the infrastructure layer, not inside the agent's code. Even if an agent's logic changes unexpectedly, the runtime blocks anything that violates policy. Agents start with zero permissions. Every escalation requires human approval. Every decision gets logged. CrewAI handles orchestration. NemoClaw handles the secure runtime. Together, organizations can run powerful autonomous agents while maintaining real control over their infrastructure and data. We've powered roughly 2 billion agentic executions over the past year and work with more than 60% of the Fortune 500. NemoClaw's infrastructure layer closes the gap between what these agents can do and what enterprises need to trust them in production.