{
  "id": 10890046,
  "title": "openai shipped always on agents and nvidia shipped the watchdog in the same week, heres what to build",
  "url": "https://urgent.news/2026/09/30/openai-shipped-always-on-agents-and-nvidia-shipped-the-watchdog-in",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-30T07:46:40.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/byteakp/openai-shipped-always-on-agents-and-nvidia-shipped-the-watchdog-in-the-same-week-heres-what-to-2d3f"
  },
  "original_language": "en",
  "account": "Two significant announcements landed this week, both of which only make sense together. On Tuesday, OpenAI launched Dots, agents that operate independently on their own cloud computer, connecting to over 4,000 apps and seeking ways to assist even without a task assigned. The day before, Nvidia unveiled the Open Agent Safety Platform, a runtime along with a separate watchdog capable of isolating a malfunctioning agent within milliseconds. One product grants agents more time and increased access, while the other exists to prevent potential mishaps.\n\nIf you construct agents, both components are now at your disposal, so let's examine the implications of an agent continuously running in the background. Unlike a normal agent with a defined shape – receiving a request, executing its task, and receiving human review for the outcome – an always-on agent lacks a finish line. There is no predetermined moment for a human to assess the agent's actions. Credentials remain accessible as the agent operates around the clock, leading to continuous costs and introducing the potential for accumulated errors. The agent learns from feedback, resulting in behavior variations from month to month compared to the tested version. In case of failures, the issues may remain unnoticed, as a wrong action occurring during the night may simply become part of the system's state.\n\nOpenAI has implemented built-in rules for autonomous actions within Dots, complemented by custom rules that enable, prohibit, or require approval for specific actions. While this appears to be appropriate, another crucial aspect requires an additional layer of oversight. Nvidia's platform comprises two layers: OpenShell, an open-source runtime software that encloses the agent, monitors its actions, and enforces policies; and Sentry, the watchdog, which operates on BlueField 4 DPUs, separate chips positioned next to the main computer. Sentry watches the agent from a position inaccessible to the agent itself. This distinction is vital, as placing the enforcement of rules on a distinct processor and within a separate process tree prevents a well-intentioned or misguided agent from altering the rules or the code responsible for enforcing them. The out-of-band watchdog maintains its own perspective of the agent's activities.\n\nNvidia emphasizes that this serves as a safety net, enforcing predetermined boundaries, rather than a universal off switch. A complaint arises from the Slashdot thread discussing this development – while Nvidia's hardware can enforce rules, it cannot evaluate whether the rules themselves are sound. Copying this concept without investing in a chip merely provides the principle: placing the referee outside the agent. In software terms, this implies the agent should never directly interact with tools; instead, every action must pass through a gate, preventing the agent process from editing the gate. The gate implementation, outlined in a few lines of Python code, includes an allow list of tools, a write budget to prevent runaway loops, an approval mechanism for destructive actions, an audit trail of decisions, and a quarantine feature to halt operations entirely when necessary. The checklist for shipping an always-on agent involves assigning scoped credentials with expiration dates, dividing actions by side effects (with reads flowing freely while writes receive a budget and destructive actions necessitate human approval), running the policy check outside the agent process and logging externally, setting spend and rate caps per day, alerting for idle or unusually busy agents, documenting potential actions when no requests are received, and periodically re-running the test suite due to feedback learning causing behavior drift. Additionally, practicing the kill switch before its necessity and measuring the time required for a full stop are recommended. The price point significantly improves this week, as OpenAI released GPT 6.1 Sol, described as near-Astra intelligence at a fifth of the cost. This affordability makes always-on agents attainable for a broader range of teams, thereby increasing the volume of deployments.",
  "summary": "two things landed this week and they only make sense together on tuesday openai launched Dots at devday, agents that run around the clock on their own cloud computer, connect to over 4,000 apps and go looking for ways to help even when nobody gave them a task. the day before, nvidia launched the Open Agent Safety Platform, a runtime plus a separate watchdog that can quarantine a misbehaving agent…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}