Securing Long-Running AI Agents: From Setup to Sandboxing

Three inflection points in AI utility

The industry progressed from the ChatGPT moment (content generation) to the DeepSeek moment (reasoning models with 10x token growth) to the current 'Claw moment' where agents autonomously execute tasks without human-in-the-loop interaction.

Agents combine models with harnesses

Long-running agents consist of reasoning models plus a 'harness'—the orchestration layer, APIs, and data sources—that enables them to independently execute code and make decisions over hours based solely on outcome-based prompts.

Neotron Ultra trained for tool use

Neotron Ultra, part of the open Neotron 3 family launching soon, was specifically trained on agent harnesses like OpenClaw and Hermes to excel at orchestration, delivering 5x speed and 30% lower inference cost than comparable models.

Skills convert libraries into agent tools

NVIDIA packaged its CUDA-X libraries as 'skills'—natural language instruction manuals that explain tool capabilities—allowing agents to understand and utilize complex software without loading entire codebases into their context window.

AIQ employs specialized multi-agent teams

The AIQ deep research skill uses a system of models where Neotron Nano routes intent, frontier models like GPT manage orchestration, and specialized Neotron 3 sub-agents handle specific disciplines like fact-gathering, critique, and synthesis.

ServiceNow autonomously resolves 90% of L1 tickets

Using NVIDIA's agent blueprint, ServiceNow deployed specialized triage and resolution agents that independently research past fixes and resolve 90% of Level 1 support tickets without human escalation.

Verified skills ensure runtime security

NVIDIA's verified skills program scans tools for vulnerabilities, evaluates cross-compatibility across different harnesses, and cryptographically signs skills so enterprises can enforce at runtime that only authorized, secure code executes.