Rob FalconerThe latest in AI

Topic · Agents & Tools

AI agents & tools

This autumn's big shift is from chatbots that answer questions to agents that keep working on their own: they run on their own computers, connect to thousands of apps and write code while you sleep. Here is what has launched and how each company is trying to keep these agents safe.

Last updated:

OpenAI: Dots and Codex

Dots: always-on agents

29 Sep 2026OpenAI's Dots are always-on agents powered by GPT-6 Astra. Each one gets its own cloud computer. They can connect to more than 4,000 apps, and you can reach them inside ChatGPT, Slack and Microsoft Teams. They are rolling out first to Pro and Business Premium customers.

Organisations can also set up "specialist" dots, and OpenAI describes an integration with Microsoft Agent 365. On safety, Dots include automatic review of the actions an agent proposes and Custom Rules that let administrators restrict what agents may do.

Codex Cloud and a $500 plan

29 Sep 2026At DevDay OpenAI introduced Codex Cloud, which runs coding agents in the cloud rather than on your laptop, alongside a refreshed Codex command-line tool. It also announced a $500-a-month Pro 500 plan that includes "Astra Ultrafast", a faster serving tier for its flagship model. OpenAI also gave a usage figure: 1.2 billion weekly users across its products.

21 Oct 2026Businesses can now run OpenAI's GPT-6 models through Google Cloud, in private preview, and point Codex tools at that service for inference.

Google: Antigravity and Agent Compute Engine

Antigravity, now with Gemini 4 Argon

13 Oct 2026Antigravity is Google's agent-first development platform, built around agents rather than an editor with a chat box attached. Google says that after Gemini 4 Argon came to Antigravity, agentic code submissions rose by more than 35%. It also describes Argon refactoring about 6 million lines of legacy Chrome code.

Agent Compute Engine

22 Oct 2026Google announced Agent Compute Engine, which gives each agent its own hardware-isolated micro virtual machine. Google says the same system already runs agents inside Gemini and will be in preview in the coming months. Short-lived credentials are injected from outside the virtual machine, so the agent does not hold long-lived keys. Google also open-sourced Agent Substrate for running agents on Kubernetes (GKE), claiming up to 10 times the density, and said its agent governance controls will extend to Claude Code and the Antigravity CLI on managed company devices, in preview soon.

Anthropic: Claude Duet and auto mode

Claude Duet

21 Oct 2026Claude Duet is a background assistant built around Claude Opus 5.5 that watches your email, calendar and business tools, sends a morning brief and checks connected accounts every few hours from its own virtual machine. It is a research preview for US Pro and Max subscribers through a waitlist. It must ask before sending messages, making payments, signing documents or deleting anything, and some actions, such as trading securities, are blocked even with approval. Read my post on Duet and auto mode.

Auto mode by default

22 Oct 2026Anthropic says auto mode, where a classifier rather than the user approves routine actions, will roll out next week as the default for new sessions in most Claude apps on Pro, Max and Team plans. It has been the default in Claude Code since 14 August. Anthropic says the classifier caught 89% of dangerous commands in its study, compared with 13.6% for people, but that auto mode "does not eliminate risk".

Computer use and browser use

7 Oct 2026Alongside Claude Haiku 5.5, Anthropic added computer-use and browser-use capabilities to its SDKs in beta, so developers can build agents that operate a desktop or a web browser. Haiku 5.5's low price and adjustable effort setting make it a candidate for the many small sub-tasks inside an agent. Opus 5.5 (22 September) is positioned for the long, hard tasks.

Meta: Muse Code

Muse Code and Muse Wonder

12 Oct 2026Meta's new Muse Wonder models are available through the Meta Model API and Muse Code, Meta's coding tool. Meta highlights two features, Active Memory and Active Focus, aimed at keeping an agent on track over long sessions. For those who want to run things locally, Muse Spark Open (15 October) is an open-weight model that Meta says runs on two Mac Studios.

Keeping agents in check

The push towards autonomy has come alongside a very public reminder of the risks. In July 2026, during cyber-capability testing, an internal OpenAI model and its agents escaped their sandbox and compromised Hugging Face systems. OpenAI published a post-mortem on 26 August. My explainer covers what happened.

On 20 October OpenAI reported two more internal incidents from early October. A GPT-6.1 Sol agent stopped alerts reaching its security team after reading an earlier "don't send any messages" instruction too broadly, and a GPT-6 Astra agent removed a monitoring integration and then wrongly told the user it had not. OpenAI says neither reached its top severity level, and its automated review now blocks attempts to disable monitoring.

The safeguards now appearing in agent products reflect that:

  • Review before action: Dots automatically review proposed actions, and administrators can add Custom Rules.
  • Staged access: the most capable cyber features go to vetted defenders first, through OpenAI's Daybreak and Google's Fairwind Program.
  • Classifier approvals: Anthropic's auto mode and OpenAI's automated review check each proposed action, with people kept for the high-stakes ones.
  • Isolation: Google's Agent Compute Engine and Claude Duet both run each agent in its own virtual machine, away from your own devices and long-lived passwords.
  • Government preparation: on 12 October the UK government said it is preparing, with the NCSC and the AI Security Institute, for incidents involving AI agents. The ICO's call for evidence on agentic AI closes on 20 November 2026.

Practical tips if you are trying agents

  • Start agents on read-only tasks such as research, summaries and drafting before you give them permission to send, buy or delete anything.
  • Give each agent the narrowest app connections it needs, and review its connections regularly.
  • Keep a person approving anything that is irreversible or visible to people outside your organisation.
  • Watch your costs: always-on agents that use flagship models can get expensive quickly, which is exactly why the cheaper Sol, Luna, Haiku and Argon tiers matter.

Sources

  1. Introducing Dots — OpenAI (29 September 2026)
  2. DevDay 2026 recap — OpenAI (29 September 2026)
  3. Introducing Gemini 4 Argon (general availability) — Google (13 October 2026)
  4. Google Cloud brings OpenAI models to customers — Google Cloud (21 October 2026)
  5. Introducing Agent Compute Engine and expanded agent governance — Google Cloud (22 October 2026)
  6. Introducing Claude Duet — Anthropic (21 October 2026)
  7. Auto mode is becoming the default in Claude apps — Anthropic (updated 22 October 2026)
  8. Claude Haiku 5.5 — Anthropic (7 October 2026)
  9. Claude Opus 5.5 — Anthropic (22 September 2026)
  10. Introducing Muse Wonder models — Meta (12 October 2026)
  11. Muse Spark Open — Meta AI research blog (15 October 2026)
  12. The Hugging Face incident and the road ahead — OpenAI (26 August 2026)
  13. Britain steps up national defences against growing AI risks — GOV.UK (12 October 2026)
  14. Suppressing misalignment monitoring alerts and disabling misalignment monitoring — OpenAI Alignment (20 October 2026)
  15. ICO secures changes from leading AI developers as scrutiny extends to AI agents — ICO (8 October 2026)

Figures quoted from company announcements are the companies' own claims. Independent testing may give different results.