# kurian-google-cloud-gemini-agent-gemini-at-work-2026-10-08

## Veille

Google Cloud blog post signed by **Thomas Kurian** (CEO of Google Cloud), published on **October 8, 2026**, adapted from his keynote at **Gemini at Work 2026**; a long text (~4,500 words) more than half of which is a catalog of customer testimonials. It announces the **Gemini agent**, a "universal agent for work": a single agent and a single API to answer questions, produce office work, create images and media, write and execute code.

## Titre Article

Welcome to Gemini at Work 2026: Introducing the Gemini agent

## Date

2026-10-08

## URL

https://cloud.google.com/blog/products/ai-machine-learning/welcome-to-gemini-at-work-2026

## Keywords

Gemini agent, Gemini Enterprise, Google Cloud, universal agent, colleague agents, sub-agents, multi-agent, multi-model orchestration, intelligent routing, spend caps, Agent Gateway, Agent Sandbox, agent identity, MCP, skill registry, tool registry, agent memory, Knowledge Catalog, Smart Storage, Borderless Lakehouse, Google Workspace, financial services, legal, TPU 8i, Claude, Gemma

## Authors

Thomas Kurian — CEO de Google Cloud ; billet du blog Google Cloud adapté de sa keynote.

## Ton

Profile: vendor keynote address, aimed at executives and CIOs. Structured as numbered announcements, with the thesis stated upfront ("work begins in the prompt window"), followed by a breakdown by domain (agent, Workspace, data, industries, governance, costs, infrastructure) and a long inventory of customers by region. Assertive, figure-heavy register, with customer results reported without methodology. The closing commitments (removing friction, removing complexity, keeping data ownership) echo the integrated-platform argument.

## Pense-betes

- **What**: the **Gemini agent**, a single agent for questions, knowledge work, media, and code, with a single API; built-in governance and cost control.
- **Colleague agents**: persistent role, own identity (email, calendar, Drive, directory), see only what is shared with them; appear under their own name in version history.
- **Memory**: session, semantic, procedural (skills the agent writes itself), episodic.
- **Model decoupled from the agent**: Gemini + **Claude** today; argument: the best model changes every few months, while context, skills, and data remain.
- **Tools and skills**: enterprise registries; connection to any internal or external **MCP** server; connectors for Salesforce, ServiceNow, Jira, Snowflake, Databricks…
- **Governance**: four questions (who is the agent, what can it do, what has it done, what must it never touch) → identity, OAuth authorizations, audit, **Agent Gateway**; identity follows the agent all the way into the execution virtual machines.
- **Costs**: per-project spend cap, agent paused once reached, one-click resume; routing between models.
- **Reported customer figures**: Bradesco document review 1 h → 5 min; Snap diagnostics 30 min → 30 s; Commerzbank 20 h → 1 h; SOMPO over 10,000 agents; U.S. armed forces, 3M users, over 100,000 agents.
- ⚠️ **Scope**: keynote announcement; preview for finance and legal, "coming soon" for public sector, healthcare, retail; no general-availability date or pricing published.

## RésuméDe400mots

On October 8, 2026, Thomas Kurian, CEO of Google Cloud, presents the Gemini agent at Gemini at Work 2026, described as a universal agent for work. It answers questions, handles knowledge work, creates images and media, writes and executes code, from a single interface and a single API. The user hands it goals rather than instructions and comes back to a finished result. Kurian cites adoption figures: nearly 500 customers each processed more than a trillion tokens over the year, nearly 80% of customers use Google Cloud's AI products, and nearly 90% of the Fortune 100 use Gemini Enterprise.

The architecture rests on five principles. The agent is unified: it converses, works autonomously, plans tasks, or reacts to events. It is accessible everywhere, including Microsoft 365 and Slack, and can operate without an interface. Its execution is persistent in the cloud, with a single memory regardless of channel. It orchestrates temporary sub-agents and colleague agents with their own identity and storage. Finally, the model is a choice distinct from the agent: Gemini chooses, for each task, between Gemini models and Anthropic's Claude models, with other models expected to follow.

Three building blocks make it useful: a tool registry, with connections to any MCP server; a skill registry, global, departmental, or personal; and four memories (session, semantic, procedural, episodic). Within Workspace, Gemini operates in Gmail, Drive, Docs, Slides, Sheets, Chat, and Calendar, as a personal assistant, through proactive delegation, or as a team-member agent with its own account.

On the data side, new skills serve data and ML engineers (PySpark code, notebooks, training, pipeline fixes) and business teams (saved operational reports, executable at no token cost). They rely on the Knowledge Catalog, which defines business terms once, on Smart Storage, which enriches unstructured data in place, and on the Borderless Lakehouse, which queries S3, Azure, and Iceberg tables without copying. Bloomberg Media reports a 63% gain in SQL query accuracy. Specializations exist in preview for financial services and legal, inheriting permissions and ethical walls from document platforms.

Governance is reduced to four questions: identity, permissions, traceability, and what the agent must never touch. Each agent has an attested identity, minimal rights, an audit log under its own name, a network sandbox, and a mandatory pass through Agent Gateway, a firewall that enforces the organization's policies. For costs, Google offers multi-model orchestration, intelligent routing, and per-project spend caps. It cites a 98% drop in price per token since 2024 and an 80% gain in price-performance from the TPU 8i.

The rest of the post lists deployments by region (Europe, Asia-Pacific, Latin America, North America) and three commitments: removing friction, eliminating complexity through an integrated stack, and keeping data with the customer.

## GrapheDeConnaissance

- Google Cloud —publie→ Gemini agent (TECHNOLOGIE, 0.98)
- Thomas Kurian —dirige→ Google Cloud (ORGANISATION, 0.97)
- Gemini agent —utilise→ Gemini (TECHNOLOGIE, 0.97)
- Gemini agent —utilise→ Claude (TECHNOLOGIE, 0.95)
- Anthropic —collabore_avec→ Google Cloud (ORGANISATION, 0.8)
- Gemini agent —utilise→ Model Context Protocol (TECHNOLOGIE, 0.95)
- Gemini agent —utilise→ Agent Gateway (TECHNOLOGIE, 0.95)
- Gemini agent —utilise→ Agent Sandbox (TECHNOLOGIE, 0.93)
- Gemini agent —utilise→ Knowledge Catalog (TECHNOLOGIE, 0.92)
- Gemini agent —s_applique_à→ Google Workspace (TECHNOLOGIE, 0.95)
- Agent Gateway —résout→ contrôle de ce que les agents ne doivent jamais toucher (AFFIRMATION, 0.9)
- Agent collègue —fait_partie_de→ Gemini agent (TECHNOLOGIE, 0.92)
- Gemini agent —réduit→ coût des agents par routage entre modèles et plafonds de dépense (AFFIRMATION, 0.88)
- Google Cloud —mesure→ baisse de 98 % du prix par jeton depuis 2024 (MESURE, 0.9)
- TPU 8i —améliore→ rapport prix-performance de 80 % par rapport à la génération précédente (MESURE, 0.9)
- Knowledge Catalog —améliore→ précision SQL de Bloomberg Media de 63 % (MESURE, 0.9)
- Google Cloud —affirme_que→ le meilleur modèle change tous les quelques mois, d'où le découplage agent/modèle (AFFIRMATION, 0.92)
- Google Cloud —collabore_avec→ Accenture (ORGANISATION, 0.88)

---
Canonical: https://www.thekb.eu/en/fiches/kurian-google-cloud-gemini-agent-gemini-at-work-2026-10-08/