Gemini now routes to Claude at work: what your company should vet

On October 8, 2026, Google Cloud unveiled, at its Gemini at Work event, a single AI agent for work that no longer just answers questions: it plans multi-day tasks, picks its own tools and, when it makes sense, hands work off to Anthropic’s Claude instead of Google’s own Gemini models (Google Cloud; CBS News). At Evicron, an AI and custom software studio based in Barcelona, what matters to us about this announcement isn’t the race between vendors — it’s the question every client is about to ask us: if Google already lets an agent act on its own inside Gmail, Docs or Drive, what should our company check before granting that same level of access?

What Google actually announced

The “Gemini agent” works from a single prompt box across Gmail, Calendar, Docs, Sheets, Slides, Drive and Chat, and also plugs into Microsoft 365 and Slack. What sets it apart from a regular AI assistant is autonomy: instead of step-by-step instructions, the user hands it a full objective and the agent decides how to get there, calling on subagents and tools along the way. Google has it in private preview for enterprise customers, with general availability expected by late October or early November, and industry-specific versions are already in preview for financial services and legal work, with healthcare, retail and government announced as next.

The detail that caught our attention most is the “coworker agent”: an agent with its own identity inside the organization — its own Workspace email address, calendar and Drive storage —, persistent memory across tasks, and access limited to only what it’s explicitly given. Today it routes work between Google’s Gemini models and Anthropic’s Claude, and Google has confirmed more providers’ models are coming later.

Why this isn’t just a product story

Until now, most businesses using generative AI did so through a chat box: someone types, the model answers, a person decides what to do with that answer. An agent with its own email address, its own calendar and the ability to act for days without direct supervision changes the underlying question. It’s no longer just “which model do we use”, but “what can this agent do on its own, and who is accountable when it gets something wrong?”. Google’s answer includes per-project spend caps, individual identity for each agent, and an audit log for every action — exactly the kind of governance any business should demand before granting this level of access, whether it comes from Google, Microsoft or an in-house build.

This ties directly into something we already covered when breaking down the EU AI Act: the more decision-making a business delegates to an automated system — hiring, customer screening, employee management — the closer it gets to the high-risk category, with its own requirements for human oversight and documentation. We go through this in detail in our guide to the AI Act’s risk tiers: an autonomous agent with access to email, calendar and internal documents isn’t automatically “high-risk”, but it does multiply the number of decisions that agent makes without anyone reviewing each one individually — which is exactly what the rules want businesses to keep in check.

What your company should check before taking this step

1. Define what the agent can see, not just what it can do. Google’s “coworker agent” model limits access to only what’s explicitly handed over; apply the same principle even if you use a different tool — no agent should inherit, by default, everything a person can see.
2. Set a cap on actions, not just monthly budget. An agent planning several days of unsupervised work can rack up mistakes — or cost — far faster than a chat where every answer passes through a person.
3. Demand an audit log for every action, especially if the agent touches customer or employee data: without that trail, there’s no way to reconstruct what happened when something goes wrong.
4. Don’t mistake “available in preview” for “ready for production.” The more advanced delegation features are still in early access; testing them in a contained environment before connecting them to production systems is plain prudence, not excess caution.
5. Start with low-risk, high-volume tasks, not processes that already rely on human judgment. It’s the same principle we applied when discussing what to automate with cheaper models like Haiku 5.5: start with the repetitive, not the delicate.

How we work on this at Evicron

An agent that operates on its own inside your email, calendar and documents isn’t something you just “switch on”: it requires deciding what access it genuinely needs, what spend and audit controls you’ll demand, and which specific task is worth delegating first. In our AI consulting for businesses work, that’s where we start — mapping access and risk — before proposing any automation, and in our applied AI projects we implement that first delegable task with the level of oversight your operation needs, whether that’s with Gemini, Claude, or both.

Bottom line

Google has turned Gemini into an agent capable of acting for days inside your email, calendar and documents, and it does so by routing some of that work to Anthropic’s Claude too. Beyond who wins the race between vendors, what matters for any business is that autonomy in these agents is no longer a future promise: it’s an option available today, in preview, for whoever wants to switch it on. The real question isn’t whether your business will use AI agents, but what access, caps and logs you’ll demand of them before letting them act alone.

Want to review what access and controls an AI agent would need before you give it a seat in your company? Get in touch: the first consultation is free and we reply within 24 hours.