Gemini Agent: Google Just Put an AI Coworker in Your Inbox

Google had a busy week. On Thursday, the company's cloud division announced the Gemini Agent — a single AI agent for work that takes a goal, goes off and does the work, and comes back with the finished result. Your next coworker might have an email address, a calendar, a Drive folder, and no pulse.

This is the third major agent launch in about a month. OpenAI shipped always-on "dots" agents in September that chase goals across apps on their own. Meta followed with Muse, a personal agent that shops, books travel and sends email for you. Now Google has planted its flag, and its version is aimed squarely at businesses — which makes it the one most of our readers should actually care about.

Here is the honest version of the story: what the Gemini Agent does, why Google is shipping its rival's models inside it, what it will likely cost, and what a sensible business owner does about it this month. Plus the rest of the week's web tech news that matters to you — because Chrome just changed the rules for your website while nobody was looking.

What the Gemini Agent actually is

Strip away the launch-day gloss and the Gemini Agent is Google's answer to a simple observation: people don't want a smarter chatbot. They want work done.

The agent plans the work, uses tools, connects to a company's business systems, and returns finished work in documents, email and developer environments. You write something like "prepare the Q4 supplier review pack," close your laptop, and the agent keeps running in the cloud — hours, even days — then delivers a finished document. It works directly inside Gmail, Docs, Sheets, Slides, Drive, Chat and Calendar, and — in a move that raised eyebrows — also reaches into Microsoft 365 and Slack.

Connectors matter here more than the model. The agent plugs into Salesforce, ServiceNow, Jira, Snowflake, BigQuery and anything that speaks Model Context Protocol, the open standard that acts as a universal plug between AI and software. An agent that cannot touch your CRM is a demo. An agent that can read your pipeline, draft the follow-ups, and log them back in is infrastructure.

Our take: the persistent part is the real product. Assistants reset every conversation; this thing keeps session, semantic, procedural and episodic memory, so it remembers business context when you switch devices or pause an assignment. That is the difference between a tool and a team member. It is also, frankly, where things can go wrong — an agent with long memory and write access to your systems needs adult supervision. More on that below.

The timing is not accidental. Google Cloud CEO Thomas Kurian said at the launch event that "work now starts in the prompt window." Translation: Google sees the prompt box replacing the inbox as the front door of office software, and it intends to own the front door.

Coworker agents: the part that changes team structure

The feature that will cause the most hallway conversation is what Google calls coworker agents. A manager describes a role — say, events coordinator or research assistant — and Gemini creates an agent with its own Workspace account: its own email address, calendar, Drive storage and a listing in the company directory.

Read that again. Your AI helper doesn't live in a side panel. It sits in your directory next to the actual humans, sends emails from its own address, books rooms on its own calendar, and stores files in its own Drive. Each agent only gets access to the information it is given, and — this is the part enterprises asked for — every action lands in an audit trail attributed to the agent rather than a person. Google gives each agent a cryptographically verified identity, a digital ID checked through cryptography, so you can tell an agent's actions apart from an employee's.

Why does this matter more than it sounds? Because it solves the identity problem that has quietly blocked agent adoption in real companies. Up until now, agents acted as you — burning your credentials, your permissions, your audit history. Nobody in compliance liked that. Giving agents their own identity, their own permissions and their own audit trail turns an agent from a liability into something a security team can approve. If you are weighing how to deploy AI agents in your business, this single design decision is the one to watch: agent identity is what makes agents deployable at scale.

The trade-off is cultural, not technical. Managers will now write job descriptions for software. Headcount planning gets weird. And the question "who approved that email?" gets a genuinely new answer: the agent did, acting within the role it was given. Companies that figure out role design for agents — what each one may do, what needs a human sign-off, who reviews the audit trail — will move faster than companies that treat them as fancy macros.

Google is starting with versions for financial services and legal work in preview, with government, healthcare and retail versions on the way. That sequencing tells you everything: regulated industries first, because that is where audit trails and identity controls are the whole ballgame.

Why Google is shipping its rival's models inside its own product

Here is the sentence that would have been unthinkable two years ago: the Gemini Agent currently runs on Google's Gemini models and Anthropic's Claude models, with more models to be added later.

Google — the company with one of the best model labs on the planet — is putting a competitor's model inside its flagship workplace product. Why? Because Google has decided the real prize isn't the model. It is the directory, the identity system, the connectors, the audit trail. Whoever controls where work happens matters more than whoever writes the smartest next token. Claude runs the task when it is the better tool for the job; Gemini runs it otherwise. A smart-routing layer picks per task, balancing quality against cost.

This is genuinely good news for buyers. It means the agent platform competes on orchestration, not lock-in. You are not betting on one model family for the next five years; you are buying a routing layer that can swap models as the leaderboard changes. For the cost-conscious, it also introduces hard project-level spending limits and committed-use savings plans — 10% off for a one-year commitment, 20% for three years. Agent workloads can spiral quietly when every task is a hundred API calls, so the spending caps are the feature finance teams should ask about first.

For our money, this validates a pattern we've been telling clients all year: the model is a commodity, the workflow is the asset. If you are building on one model provider with no abstraction layer, the October 2026 news cycle should make you nervous.

The agent race, October 2026: a buyer's comparison

Three launches, one month, three different theories of what an AI agent is for. Here is how they line up for a business owner:

Our judgment call: if you are buying for a company, Google's is the only one of the three designed for procurement. Identity, audit, and spending controls are the three things that turn an agent pilot into a deployable system — and Google shipped all three on day one. Meta Muse is the consumer product to watch for how customers will expect to interact with your business. OpenAI's dots sit somewhere in between.

We have covered the broader pattern in our guide to deploying AI agents in your business and our look at persistent agents on autopilot — both worth a read before you sign anything, because the governance questions haven't changed even though the products have.

What will go wrong: the failure modes to plan for

We work with agents in production, so let us say the quiet part out loud. Everything in the launch announcement is the happy path. Here is what the first year of coworker agents will actually look like inside real companies, and what to do about each one.

Confident mistakes at machine speed. An agent with write access to your CRM and your inbox will eventually do the wrong thing quickly and at scale — send the follow-up to the wrong segment, update the wrong record, book the wrong meeting. Humans make these mistakes too, but humans make them one at a time. The fix is boring and non-negotiable: approval gates on every consequential action. Drafts are fine to automate; sends, writes and payments get a human click. Google says the agent asks for approval before consequential actions; make sure your policy actually enforces that instead of waving it through.

Stale memory. These agents remember — session, semantic, procedural, episodic. That is the selling point, and it is also a landmine. An agent acting in October on supplier terms it learned in March, or on a pricing sheet that changed last week, produces work that looks right and is wrong. The fix: treat agent memory like a database with a refresh policy, not like a colleague's judgment. Anything time-sensitive gets re-verified against the source system before it leaves the building.

Permission sprawl. Connectors to Salesforce, Jira, Snowflake and everything with an MCP plug sound wonderful until you realise each connector is a permission surface. Give every coworker agent the keys to the kingdom and you have built the world's most efficient insider threat. The principle is least privilege per agent: the research agent reads the web and your docs, but it does not touch the CRM. The sales agent touches the CRM, but it does not touch finance. Role design is access design.

Invoice shock. A single "prepare the Q4 review pack" task can chain dozens of tool calls and model invocations, each metered. Multiply by a team of agents running daily and the bill stops looking like software and starts looking like payroll. Set the project-level spending limits Google provides on day one, review usage weekly for the first quarter, and treat the committed-use discounts as a year-two conversation, not a launch-day decision.

Accountability theatre. The audit trail is the most important feature in this whole launch — and the easiest to waste. An audit log nobody reads is decoration. Somebody in your company needs the job of reviewing agent actions: a weekly fifteen-minute scan of what the agents did, what they touched, what got flagged. Without that human in the loop, you have not bought governance. You have bought the appearance of it.

None of this is a reason to sit out. It is a reason to pilot with your eyes open. Agents fail in predictable ways, which means the failures are manageable — if you design for them up front instead of discovering them in a post-mortem.

What it will cost: the numbers so far

Google gave no pricing or general availability date for the Gemini Agent itself, which is normal for a preview-stage enterprise launch. But there is enough public pricing around it to sketch the bill. All figures below exclude local taxes, and Pakistani buyers should budget for card-processing quirks on international SaaS billing.

A ten-person team on Business Standard pays roughly $140 a month — about PKR 39,000 — and already has Gemini inside every document. The cheapest meaningful upgrade path, moving from Starter to Standard, is $7 a head. That is the number to remember: your agent journey probably starts as a $7 decision, not a rip-and-replace.

Two cost warnings from the studio. First, model routing is a meter, and meters run. Agents that chain dozens of tool calls per task can produce surprising invoices; set the project-level spending limits before your first pilot, not after. Second, committed-use discounts of 10–20% look tempting, but do not pre-commit to a platform that is still in preview. Buy monthly, run a real workload for a quarter, then talk about commitments.

For teams in Pakistan, the honest arithmetic: $30 a seat is serious money — around PKR 8,400 per person per month — which is why the Workspace-bundled route matters. If Gemini is already inside the $14 seat you pay for, the marginal cost of experimenting is zero. Start there.

Do a rough break-even before you pilot. Take one workflow — say, weekly reporting that eats five hours of an operations person's week. At a PKR 150,000 monthly salary, those five hours cost you roughly PKR 17,000 a month in loaded time. If an agent on a $14 seat (about PKR 3,900) absorbs even half of it, the maths clears by a factor of two. That is the kind of arithmetic that turns a pilot into a budget line — but only if you measured the hours before you started, not after.

What business owners should do this month

Not adopt it. Not yet. Here is the four-step move we recommend to clients, in order:

1. Audit what you already pay for. Most ten-person companies we talk to pay for three overlapping AI subscriptions — a ChatGPT team plan, a Copilot seat or two, some Gemini add-on nobody remembers buying — and use none of them properly. Kill the duplicates before you add a fourth. That audit usually pays for itself in a month.

2. Pick one repeatable workflow and run a four-week pilot. Inbox triage. Meeting notes to action items. Lead follow-up drafting. Pick the one that eats the most staff hours, give one platform a single workflow, and measure: hours saved, errors caught, time to output. If you cannot measure it, you cannot justify it — and "it feels faster" is how AI budgets die.

3. Learn the governance controls before you scale. Agent identity, audit trails, spending limits, human approval for consequential actions. Read our AI agent safety playbook if you haven't. The companies that get burned by agents in 2027 will be the ones that skipped this step in 2026.

4. Wait for pricing to settle before committing a department. The 10–20% committed-use discounts are bait. Previews become products, prices move, competitors respond. Run monthly, decide quarterly.

And the one thing to start this week: check your Workspace admin console. Agent features land there silently, and the admin who discovers them first gets to set the policy. The admin who discovers them last gets to clean up the mess.

This week in web tech: the rest of the news that affects your business

The Gemini Agent was the headline, but three other stories from the past week land closer to home for anyone running a website.

Chrome is going HTTPS-by-default this month

Starting with Chrome 154, released this month, the browser enables "Always Use Secure Connections" for everyone. Chrome will attempt every connection over HTTPS first, and it will ask for your permission before loading any public site that doesn't support it.

For most sites this changes nothing — over 95% of Chrome page loads already use HTTPS. But if any part of your web presence still serves over plain HTTP — a legacy subdomain, an old landing page, a staging link that leaked into Google — your visitors will now hit a warning screen. That is a trust killer and a conversion killer.

Fix it this week: audit every property you own for HTTP connections, put certificates on everything public, and test with the setting switched on in Chrome today. If your site is on our stack, this is a service we handle routinely — but any host can issue a free certificate in minutes. There is no excuse left for HTTP in October 2026.

Microsoft bets $2,500 on the AI PC

Microsoft unveiled the Surface Ultra this week — a laptop starting above $2,500 (roughly PKR 700,000) built around Nvidia's RTX Spark AI chip, with Satya Nadella and Jensen Huang sharing a stage to sell a "hybrid intelligence" future: AI that runs on your device instead of in a data centre.

Our read: the hardware is not the story. The story is that every platform vendor now agrees AI workloads are splitting — some in the cloud, some on the device — and software should stop caring which. For businesses, that means the next generation of your team's laptops will run agents locally: faster, offline-capable, and cheaper per task than cloud calls. If you refresh hardware on a three-year cycle, the machines you buy in 2027 will be agent-native. Plan your software budget accordingly.

Kagi gives up on Orion for Windows and Linux — and open-sources it

Smaller story, bigger lesson. Kagi announced it is ending development of its Orion browser for Windows and Linux and handing the code to the community, focusing instead on macOS and iOS. Orion is built on Apple's WebKit rather than Chromium, and porting WebKit to other platforms turned out to be a maintenance nightmare for a small team.

The lesson for founders: platform bets are strategy, not preference. Kagi chose focus over sprawl, and that is the right call for a small team — spreading thin across platforms is how good products die slowly. If you are deciding where to build, read our build vs buy guide before you commit headcount to a second platform.

Frequently asked questions

What is the Gemini Agent that Google launched in October 2026?

The Gemini Agent is Google Cloud's persistent AI agent for work, announced on 8 October 2026 at its Gemini at Work event. It takes a goal, plans the work, uses tools across your business systems, and returns finished work in documents, email and developer environments. It keeps running in the cloud after you close your laptop, so a single task can stretch across hours or days.

What are coworker agents?

Coworker agents are Gemini agents provisioned as digital team members. A manager describes a role — events coordinator, research assistant, inbox triage — and Google creates an agent with its own Workspace account: an email address, a calendar, Drive storage and a listing in the company directory. Each agent only accesses the information it is given, carries a cryptographically verified identity, and every action is recorded in an audit trail attributed to the agent rather than a person.

Does the Gemini Agent use Google's models only?

No — and this is the surprising part. The agent currently runs across Google's Gemini models and Anthropic's Claude models, with more models to be added later. A smart-routing feature picks the best model per task for quality and cost. For buyers, that means Google is competing on being the best orchestration layer for work, not on locking you into one model.

How much does the Gemini Agent cost?

Google announced no specific pricing or general availability date for the agent itself. For context, Gemini is already bundled into Google Workspace seats: Business Standard runs about $14 per user per month on annual billing and includes Gemini in Docs, Sheets and Slides, while Business Starter at $7 does not. Google's heavier Gemini Enterprise tier has historically been priced around $30 per user per month. Expect the full agent platform to be priced at or above that Enterprise tier.

How does the Gemini Agent compare with Meta Muse and OpenAI's dots?

All three launched within weeks of each other in September–October 2026, but they aim at different buyers. OpenAI's dots are always-on agents that chase goals across consumer apps; Meta Muse is a personal agent that shops, books travel and sends email for individuals. Google's Gemini Agent is the enterprise play: it lives inside company systems, gets a corporate identity, answers to an audit trail, and works across Workspace, Microsoft 365 and Slack. If you run a business, Google's is the one built for you.

Is the Gemini Agent available in Pakistan?

Availability will follow Google's standard Workspace rollout pattern. Business Starter and Standard plans with Gemini are broadly available in Pakistan, and the agent platform builds on those same systems. If your team already runs on Workspace, you will likely see the agent features appear in your admin console before any formal launch announcement for your region. The vertical editions for financial services and legal work are in preview first.

What should a small business do about AI agents in October 2026?

Start by auditing where your team already pays for AI — most businesses we meet pay for three overlapping subscriptions and use none of them properly. Then pick one repeatable workflow (inbox triage, meeting notes, lead follow-up) and run a four-week pilot with one platform. Do not hire agents as staff yet; buy one platform, learn its governance controls, and wait for pricing to settle before committing a whole department.

The bottom line

October 2026 will be remembered as the month AI agents grew up. OpenAI proved agents could chase goals across apps. Meta proved they could live in your pocket. Google proved they could sit in your org chart — with an email address, an audit trail, and somebody else's model doing half the work.

That last one is the version that changes how businesses run. An agent with identity, permissions and a paper trail is no longer a demo; it is staff. And staff needs management: role design, spending limits, review rhythms. The companies that treat agents like employees — with onboarding, boundaries and oversight — will get the productivity. The ones that treat them like software will get the incidents.

If your team is figuring out where agents fit — or you just want someone to audit what you are already paying for — talk to us. We build the workflows, set the guardrails, and measure the results. The agents are getting good. The question is whether your business is ready for them.

---

Sources: Reuters on the Gemini Agent launch · Decrypt on coworker agents · Zubiqo on smart routing and savings plans · Memeburn on Workspace and Claude pricing · Infosecurity Magazine on Chrome 154 HTTPS-by-default · Malay Mail on the Surface Ultra launch