Google's work agent can have its own inbox. Start with the four questions.
On 8 October Google described a coworker agent with its own email. The useful part of the keynote is the governance. The cheap model is how the scope creeps.

On 8 October 2026 Thomas Kurian published the keynote from Gemini at Work. The sentence to underline is his, not a paraphrase from a recap: you give it objectives, not instructions. You delegate an outcome and come back to finished work. The agent that is supposed to do that can sit in Gmail, Drive, Docs, Slack, and Microsoft 365, and a coworker version of it can have its own Workspace account: an email address, a calendar, Drive, and a line in the company directory.
A day earlier, Anthropic's list prices for Haiku 5.5 went live. On prompts under 100,000 tokens the rate is $0.10 per million input tokens and $0.50 per million output tokens. Haiku 4.5, on the same rate card, is $1 and $5. Past 100,000 tokens Haiku 5.5 is $0.50 and $2.50, half the older small model. A tenth of the price on the common case, plus a product that wants a standing inbox, is how "try it on one workflow" becomes "it answers as the company."
8 Oct
Gemini agent described in the Cloud keynote
One agent for questions, documents, images, and code. A coworker can hold its own Workspace identity.
$0.10
Haiku 5.5 input, per million tokens, under 100,000
Output is $0.50. Haiku 4.5 on the same card is $1 in and $5 out.
4
Governance questions Google put in the keynote
Who it is, what it may do, where the audit is, and what it must never touch.
What the keynote is selling
Read the post as a design, then decide which pieces you would actually turn on. Google describes one agent you reach from the web, a phone, a desktop, the command line, Workspace, Microsoft 365, or Slack, including a headless API with no screen of its own. It keeps one memory across those surfaces. Close the laptop and a long job keeps running. It can spin up temporary sub-agents, each with an identity, and it can also act as a coworker: a standing role that lasts across days, with its own storage and only the context someone shared with it.
The model underneath is a separate choice. Today that choice includes the Gemini family and Claude. Google's line is that the best model for the task is not always the largest one, and that the leading model changes every few months, so the context and the skills should stay put when the model does. That split is correct. We have been saying a version of it since the cost piece in June. The keynote's version arrives attached to a directory listing.
Four kinds of memory sit under the pitch. Session memory for the task in front of it, even when the task runs for days. Semantic memory, a knowledge base it builds as it reads and talks. Procedural memory for how a job gets done, including skills it writes for itself. Episodic memory of what it has already done. The post says the agent onboards itself the way a new hire would. A new hire has a manager, a start date, and a list of systems they do not get on week one. An agent that onboards itself has whatever was in the thread.
The post does not give a date when every Workspace customer can create one of these. Treat the inbox as a permission you have not granted. Customer counts in the same article — thousands of agents, hours saved, a threefold conversion — are the vendor's stories. They are not your pilot.

Steal the four questions, refuse the default
The governance section is the part worth keeping even if you never buy this product. Google asks four things before an autonomous agent is allowed to roam:
- Who is the agent, and what identity does it have?
- What is it allowed to do?
- What did it do, and where can you see it?
- What should it never touch?
Their answers are an identity that is attested and logged in the agent's name, permissions a security admin approves, an audit trail that does not hide inside a person's account, a sandbox with its own network boundary, and a gateway that can enforce a rule like "agents may not open Need-to-Know documents" for every agent at once. A twenty-person company does not need that gateway to answer the same four questions on a page. The production guide and the Hugging Face post were already that page. The news is that the keynote leads with the questions. The news is also the sentence that comes first: objectives, not instructions.
| A universal agent | A job | |
|---|---|---|
| Goal | Handle my work | When this happens, do this, then stop |
| Identity | A standing inbox, calendar, and directory line | A run, then nothing left behind |
| Memory | Learns you, the team, and the history | The task in front of it |
| Model | Routed, which is the right idea | Routed, with a dollar cap |
| Failure | It was helpful in the wrong system | It stopped |
Renting the loop is still the right call when the job ends in a file. A coworker with an email address is a different product. It can be @mentioned in a chat, tagged in a document comment, and show up in version history under its own name. Google says it acts under that identity and sees only what you share. That is the right shape after a job has been boring for a month. It is the wrong shape for the pilot. An email address is a standing permission to talk to the rest of the company, including anyone who forwards the thread.
The cheap model is how the scope creeps
Haiku 5.5 is a good tool. Summaries, classification, a first pass over a pile of tickets. The June argument still holds: stop paying frontier prices for grocery math. The new failure runs the other way. The grocery-math model is now cheap enough, and good enough, that someone will point the universal agent at the inbox because the bill looks small. A tenth of last month's small-model rate does not make a company identity a bargain. It makes the mistake affordable.
Google's own cost section admits the tension. They say per-token prices have dropped sharply since 2024 and that volume has exploded, so running every simple loop through a premium model breaks the budget. The control they announce next to that sentence is the one to copy: a hard spend cap on a project, enforced in the billing console, that pauses the agent when the cap trips. You can resume with a click. Copy the cap. Then remember what it is. A pause on spend stops a bad week of tokens. It does not unwind a bad email that already left.
Smart routing belongs in the same paragraph. Let a router pick a small model for the easy step. Put the cap on the project, not on your memory of the invoice. Name the person who gets the pause alert. If nobody is named, the cap is a setting.
One job you can say in a sentence
a trigger, an output, a stop
|
v
A run with its own name in the log
no company email yet
a dollar cap before the second week
|
v
A standing inbox
email, calendar, directory
only after the job has been boring
What to do when the demo lands in a meeting
Write this down before anyone creates the coworker. It fits on the same page as the one-sentence scoping test.
- Name one job. "Triage the shared support inbox and draft replies a person sends" is a job. "Be the operations coordinator" is a role, and a role is how you get the directory listing on day one.
- No company email until that job has run with a human on every write. Drafts are fine. Sending, commenting as itself, and accepting calendar invites are the promotion.
- Give the run an identity that is not an employee. Logs in the agent's name. If a person has to lend their login, you do not have an audit. You have a shared password with better marketing. The shadow AI problem is people pasting work into tools you do not know about. A directory agent is the official version of the same leak if its memory is "everything we have ever said."
- Set the dollar cap before the second week. Include sandbox time if the product bills it separately. The person on the alert should be someone who can pause the job without opening a ticket.
- Decide what memory is allowed to keep. Session notes for the task are reasonable. A standing knowledge base of every document it has read is a retention policy. Write the policy, or do not turn that memory on.
- Apply the allowlist to sub-agents. A roster of temporary helpers is still your blast radius. If the parent can create them, the list of systems they may touch is the parent's list, not a fresh negotiation.
Give it a job and a cap. The inbox is a promotion it has not earned.
The keynote is a serious design. Identity, an audit in the agent's name, a gateway for the things it must never touch, and a spend cap that actually pauses the work: those are the pieces a buyer should demand from whoever hosts the loop, Google included. The piece to decline, until a single job has been dull for a while, is the coworker with the company email and a memory that hired itself.


