Last week our featured story was OpenAI's dots, the always-on agent that turned up inside a $100 ChatGPT plan with its own cloud computer. Back on 1 September we led with OpenAI and Anthropic putting agents behind users' logins, and the questions we set then were about authority. On 8 October Google Cloud joined in and went a step further than either. Its agent does not borrow your login. It gets its own.
At Gemini at Work 2026, Thomas Kurian, CEO of Google Cloud, announced the Gemini agent: one persistent agent, run in Google's cloud.
Plans the work, uses skills and tools, connects to your systems, and brings back something finished.
The piece that matters to a project team is what Google calls a coworker agent. You describe a role, and Gemini creates an agent with its own Workspace account: an email address on an @agents.company.com domain, a calendar, Drive storage and a listing in the company directory. Colleagues add it to a Chat space or @mention it.
Google's own description is that the agent can work "on behalf of a group of people, like a project manager within a team". That line should prompt a pause rather than a purchase. A project manager carries accountability under a contract. An agent carries a log.
What Google actually shipped
The agent runs in the cloud with one set of memories whichever device you use, so Google says work "that takes hours or days keeps running after you close your laptop". It connects to Microsoft 365, Slack, Jira, Confluence, Salesforce, ServiceNow and any Model Context Protocol server, and Kurian says it routes each job across the Gemini family and Anthropic's Claude models.
TechCrunch's Sarah Perez reports a 'tasks inbox' where users watch its reasoning and its delegation to sub-agents.
Businesses come first. Perez reports Sundar Pichai telling the event that starting with enterprises lets Google work through the "harder problems around security, scale, and performance" before any consumer release.
Decrypt's Jose Antonio Lanz notes that Google published no pricing and no general availability date, so the practical questions for a UK firm still sit with your reseller.
Its own name in version history
Inside Workspace, Google says a coworker agent "acts under its own identity rather than yours, and it sees only what you share with it". Tag it in a document comment, and it suggests an edit, replies in the thread and appears under its own name in version history.
On a live project, the risks become more concrete:
Someone tags the agent in the RFI tracker and asks it to draft a reply to the contractor’s query about the slab edge detail.
The draft appears under the agent’s name in the version history of a document visible to the client.
Sign-off procedures need to be established. Under a JCT or NEC contract, a reply comes from a named role, such as the contract administrator or project manager. The agent is not a party to the contract.
One rule should be in place before the account is created: the agent’s name must never be the last entry in the version history before a document is issued.
Govern it and afford it
The most useful line in the keynote splits the governance case in two.
Two factors determine whether an enterprise agent program succeeds or stalls: whether you can govern it, and whether you can afford it.
On governing, every agent gets a cryptographically attested identity with least-privilege permissions, stamped into its logs and any virtual machine it spins up. Every action goes to an audit trail "attributed to the agent rather than to a person". All traffic passes through Agent Gateway, which Google calls an AI network firewall, applying a policy such as "agents may not open documents classified Need to Know" to every agent at once. On affording, you set a hard spend cap per project. If it trips, that project's agent pauses until someone resumes it with a click.
Both are real controls, and both move a question onto the project team. An audit log that names the agent satisfies IT. When something goes wrong, the client will still ask which human approved it, and the log will not say. A cap that pauses an agent halfway through a Friday valuation is a programme risk, and the programme owner should set it.
The proof points, with caveats
Google offers two customer figures. Hitachi, it says, has field technicians who have built agents for more than 200 use cases, including image-recognition agents that verify photos of every tool and wire before and after a job, which Hitachi credits with productivity gains of up to 30%.
Tata Steel is said to have deployed more than 300 specialised agents in nine months and cut customer complaint turnaround by 50%. Both are vendor-relayed, the 30% is an 'up to' and no methodology is published.
The Hitachi example is still the one closest to a UK site. Before-and-after photo checks of tools and cabling are what snagging and handover already generate. If a coworker agent with its own login first appears on one of your projects, that is the likeliest door.
Takeaway
The questions from 1 September still stand. Google's launch adds two more: who sponsors the agent's account, as a line manager would a new starter, and what its name in a log means when a document leaves the firm. Pichai's business-first logic cuts both ways: your firm is the test bed.
Decide whether an agent may hold its own account on a project tenant at all, and if so, name a human sponsor in its directory entry.
Write the issue rule now: the agent drafts, a named person issues, and nothing reaches a client, contractor or bidder under an agent identity.
Put the per-project spend cap on the risk register.
Ask your reseller for pricing, UK availability and whether coworker agents sit inside your existing Workspace data processing terms.
We will keep tracking how agent identity, audit trails and sign-off rules settle across Google, Microsoft and OpenAI as the first UK project teams switch these on. The weekly Project Flux newsletter is where we report it: sign up here.
Links and Stuff
All content reflects our personal views and is not intended as professional advice or to represent any organisation.
