Hi subscribers!

Welcome to the third edition of The Frequency.

I had a whole opening here about Victor Wembanyama losing sleep over the Knicks winning the Finals.

But as I was about to schedule this, OpenAI released Dots, an always-on agent with its own cloud computer. They can carry work across ChatGPT, Slack, and Teams, which is almost comically well timed for an issue about AI agents leaving the laptop.

So, the Knicks (and I guess Wemby) can wait.

In this issue, we’re looking at where “agents” are running in the workforce:

  • Personal computer-use agents showing up in work systems

  • Why chat is becoming the front door for always-on agents

  • Why agent harnesses may be expanding from laptop to cloud

Before diving in: one quick plug for my 🇬🇧 London-area 🇬🇧 subscribers. On November 4, we’re bringing Inference: London to Canary Wharf for a practical day on agents, Claude, governance, and AI legislation. Anthropic’s cybersecurity community (incl. our very own Ed Merrett) will then take over for an Impact Lab hackathon.

Bring a laptop and hopefully leave with something more useful than a tote bag.

Computer-use agents would like a work login 👀

Last issue, I wrote about Grok Bot and Meta AI for Mac. The bigger question now is whether employees will connect personal agents like these to work. Grok Bot can stay logged into apps through a cloud computer, while Meta’s Mac app can use on-screen context and type into other applications. Both make it easy for a personal assistant to become an unmanaged work agent.

What you should care about: A personal AI assistant becomes an unmanaged work agent the moment someone connects it to Slack, Gmail, Salesforce, or another company tool. It may have access to what is on screen, active browser sessions, files, and SaaS permissions, often through an employee’s personal account.

Treat that work connection like an onboarding event, not an app download. Before one touches company systems, check who owns the account, how it authenticates, what it can see and do, which connectors it uses, what gets logged and retained, and when human approval is required.

Chat is becoming the front door for always-on agents

The old distinction between “ask an AI something” and “hand an AI some work” is fading. Claude is folding Chat and Cowork together, while Microsoft is building Copilot Home around the same premise. Then, as I was about to schedule this, OpenAI released Dots: always-on agents with their own cloud computers, built to carry work across ChatGPT, Slack, and Teams.

All three are moving toward the same experience: tell the assistant what you need, then let it determine whether the work happens in a chat or continues in the background. The product looks simpler. The permissions, data access, and execution behind it do not.

Anthropic began merging Claude Chat and Cowork on September 16, starting with Pro and Max users, with at least 30 days’ notice promised for Enterprise admins. Microsoft announced Copilot Home on September 25, alongside Code and Autopilot. Home and Code will begin rolling out through Microsoft’s Frontier program in the coming weeks; Autopilot, which runs in the cloud and can keep working after the user steps away, remains in private preview. On September 29, OpenAI announced Dots, always-on agents with their own cloud computers. Dots is rolling out first to eligible Pro and Business Premium users, while Enterprise access is an admin-enabled beta.

What you should care about: A routine chat can now turn into background work in a cloud environment, with access to company data or another business system. Visibility and guardrails need to follow the activity, not the tab where it began.

Your AI agent is leaving the laptop

Most agent-security conversations start with the employee’s laptop. The agent reads local files, opens a browser, runs a tool, and asks for approval. While work is still happening on the laptop, the software coordinating that work is increasingly expanding to the cloud.

That software is called the harness. It gathers context, keeps track of the task, chooses tools, asks for approvals, and manages execution. Claude Cowork now runs its agent loop and code execution in the cloud by default, while Claude Desktop can still reach local files and applications. OpenAI Codex supports a similar mix of local and cloud environments.

This is great for anyone who has ever closed their laptop halfway through a long-running task. Buuuut it’s slightly more awkward for the security team trying to understand what happened afterward.

Here’s the deal: A task can now start in Slack, build context in a cloud-hosted agent, access a local file, use a connector, and change an enterprise system. Your endpoint may show the file update. Your identity system may show the connector authorization. The AI provider may show the prompt and tool call. None of them tells the whole story alone.

What you should care about: Endpoint controls still matter, but endpoint visibility is no longer the full answer. For agent workflows that can read, write, send, or change data, you need evidence across the provider, endpoint, connectors, and destination system. You should be able to trace who started the task, what context it used, what access it had, what was approved, and what changed downstream.

WHAT WE’RE LISTENING TO RIGHT NOW 🎙

As you’ve hopefully caught on by now, I curate a playlist each month from a different team here at Harmonic. This month, we are featuring our Product Discovery Team. This is the team that decides what to build and why it matters. Listen to their September playlist below:

Finally, I’m always looking for feedback. See something you liked, or something you didn’t? Want to see something in the next issue? Got a cool AI tool you want to showcase? Hit reply and let me know.

Catch you next time,
Tyler

❝

Know a security lead who'd find this useful? Forward it their way.

And if this was forwarded to you, subscribe at harmonic.security/frequency