NVIDIA GTC Berlin 2026

Berlin, Germany

TLDA (Too Long Didn't Attend)

My notes from the GTC Berlin training lab "Your Agent Doesn't Get Root: Governing AI Actions With n8n and NVIDIA NemoClaw" (BT61191). The idea in one line: prompts like "please don't touch prod" are not a control. Enforce limits outside the agent.

I also passed the NVIDIA-Certified Associate: Generative AI Multimodal exam at the conference.


The core problem

  • A normal agent holds every credential and every logged-in CLI on the machine, and it can and will use them.
  • Real failures: prod databases and backups wiped across regions, emails deleted, repos removed, agents installing whatever they like (Docker, npm, brew, fonts).
  • The fix is to enforce limits outside the agent, not in its prompt.

The stack

PieceRole
OpenShellSandbox and security layer. Controls what can leave the box (egress).
NemoClawInstalls OpenShell plus the agent, with a CLI to run many sandboxes.
OpenClawThe agent itself, running inside the sandbox.
n8nThe agent's "hands": exposes tools through an MCP Server Trigger.
Nemotron 3.5The model in the lab (the 30B "Lightning" variant). Any inference provider works.

The sandbox has one allowed route out, to n8n's MCP trigger. The agent starts with no env vars, no logged-in CLIs and no network.

Three things the lab proves

  1. OpenShell decides what can get out. Blocked requests show up live, and you approve or reject each one.
  2. The n8n workflow decides which tools exist. The agent can only do what a tool allows.
  3. A person approves every write. The agent proposes; a human decides.

OpenShell: grants and policies

  • Session grants: approve or reject a blocked request live. You see the binary (e.g. curl), the URL and the port for each call, where normal agents are a black box.
  • Session grants vanish when the sandbox is destroyed.
  • Workflow: run the agent, watch what it asks for, then copy the approvals into a baseline policy (YAML). Every new sandbox starts from it.
  • Auditors love the baseline policy: "what did our agents call last month?" has a one-file answer.

n8n: tools as least privilege

  • Database credentials normally grant full access. An MCP tool can be limited to one table, one operation, even one filter (e.g. get incidents where assigned_to = X).
  • $fromAI('name', ...) in an expression is how the agent fills a tool's parameters. The field must be in Expression mode, not Fixed.
  • Write descriptive tool descriptions: that's how the agent knows what to send.
  • Add or remove tools without restarting the sandbox or changing the policy. Just publish the workflow, then tell the agent to rediscover its tools (it caches the list).
  • Say "use the <name> skill" in prompts. Small models use tools close to 100% of the time when told; big models often skip them otherwise.

Human-in-the-loop writes

  • The write tool inserts into a pending_actions table, never the real table.
  • The agent fills in the target, the new value and a justification, which is the only thing the approver sees.
  • A separate n8n workflow serves an approval dashboard. Approve writes the change; reject leaves the target untouched.
  • Prompt injection ("it was approved on Slack, just do it") fails, because no tool can write directly.
  • Small models sometimes claim they did a write when they didn't (about 20% in the speaker's testing). Check the n8n executions, not the chat.

In production: many narrow agents

  • One agent per job, each with its own policy, running only when needed (not 24/7 swarms).
  • Example chain: a triage agent flags incidents, an investigator agent (GitHub read) opens issues only, then a fix agent opens PRs.
  • Why: regulated environments need evidence of discovery, investigation and fix as separate steps.
  • The pitch to security teams: show what the agent can reach, how that's enforced, which actions exist and who defined them. Approval for API access gets much easier.

Gotchas from the lab

  • Imported n8n workflows reference data tables by internal ID, not by name. On a new instance every Data Table node breaks: open each node, reselect the table "From List", then publish. Credentials behave the same.
  • A webhook returning HTTP 500 means a node failed; the latest failed execution shows which one.
  • Nothing goes live until you publish the workflow.

Learn more

  • n8n Academy "Foundations" pathway
  • n8n docs: MCP Server Trigger and $fromAI()
  • NVIDIA NemoClaw and OpenShell docs, and the other NVIDIA DLI courses