AI News Roundup — October 2, 2026

This edition is about agents that keep running after you walk away, and about the guardrails being built around them. OpenAI used DevDay to launch persistent agents and a much cheaper model to drive them, Nvidia put hardware underneath agent containment, and the largest AI developers signed a voluntary safety pledge at the White House while a federal regulator opened its own inquiry.

OpenAI launched Dots, agents that keep working after the chat window closes. According to VentureBeat, Dots are built on GPT-6 Astra and are designed to monitor projects, use software and bring completed work back for human approval. Heise reports that each Dot gets its own cloud computer with a browser, can reach more than 4,000 apps through plugins, and can hand work to sub-agents. ChatGPT Pro and Business Premium customers get one Dot at no extra charge; Enterprise, Edu and Healthcare get it as a beta their administrators must enable, and Pro users in the EEA, Switzerland and the UK are excluded at launch. Organisations decide through custom rules which actions a Dot may take alone, which need approval and which are forbidden.

Alongside it, OpenAI released GPT-6.1 Sol at a fifth of Astra’s token prices. The Next Web reports Sol costs $2 per million input tokens and $10 per million output tokens, with cached input at $0.10, and is aimed at coding, computer use and business workflows in ChatGPT, Codex and the API. MarkTechPost says it has a context window of about 1.05 million tokens and matches Astra on the DeepSWE v1.1 coding benchmark. For context on the flagship it undercuts, see our coverage of the GPT-6 Astra launch.

Nvidia launched an Open Agent Safety Platform to contain agents that go off-script. In its announcement, Nvidia pairs its open-source OpenShell runtime, which sets a boundary around what an agent can reach, with Sentry, a reference design for an out-of-band watchdog on BlueField-4 data processing units that can quarantine an agent in milliseconds. More than 100 organisations are collaborating on it, including Anthropic, Microsoft, Cisco, Salesforce and SAP, according to The Daily Star. “AI’s extraordinary potential for society will only be realized if we solve AI safety,” said Nvidia CEO Jensen Huang. The launch lands days after the rogue-agent incident we reported last week.

AI leaders signed a voluntary safety accord at the White House, and the FTC is investigating anyway. Al Jazeera reports that Meta, Nvidia, Google, OpenAI, xAI and Anthropic signed the White House Accord on Super Intelligence, which commits them to four layers of control: internal controls, an internal oversight team, independent outside auditors and a board committee to review results. It carries no penalties and does not require audit results to be published. Separately, CNBC reports that the Federal Trade Commission is investigating OpenAI, Anthropic and other AI companies over product risks, and Reason, citing the New York Post, says the agency plans civil investigative demands and executive testimony under its existing consumer-protection powers.

Our take. Always-on agents with their own computers, cheaper models to run them, and containment layers outside the model all point the same way: the hard engineering is now permissions, monitoring and audit trails, not prompting. The accord’s four layers read like a checklist any business deploying agents should already have, whether or not a regulator asks. Spaqi builds custom AI agents and the cloud and DevOps plumbing around them, including the approval rules and logging that make them safe to leave running. If you are weighing persistent agents, talk to us.

Sources: VentureBeat, Heise, The Next Web, MarkTechPost, NVIDIA Newsroom, The Daily Star, Al Jazeera, CNBC, Reason.