On October 8, Google Cloud unveiled the Gemini agent at Gemini at Work 2026. A "universal agent for work" merging Q&A, knowledge work, content creation, and coding into one surface. CEO Thomas Kurian's line sums up the direction: "Give it objectives, not instructions. Delegate an outcome and come back to finished work."

Enterprise AI is moving from standalone chatbots to agent platforms wired into business systems. This post covers what arrived and what to check before adopting.

What arrived: one window, everything behind it

User: delegates an objective ("wrap up this quarter's report")
  ↓ one prompt box, one API
Gemini agent:
  ├─ plans work → picks skills and tools → connects systems → returns finished output
  ├─ works inside documents, inbox, and dev environments you already use
  ├─ spins up subagents for parallel or sequential long jobs
  └─ model choice, cost controls, and governance built in

Reachable from:
  web, mobile, desktop, CLI, Workspace, Microsoft 365, Slack,
  headless embedding in third-party apps
  invoked via @Gemini mentions, email, sharing, schedules, events

The bet is omnipresence. For Workspace shops and for companies staying on Microsoft 365 and Slack alike. For developers the key line is "the agent is the product and the model is a routing decision": each job runs on whatever model fits.

Coworker agents and memory: AI registered like people

The heaviest announcement is the coworker agent.

Coworker agent = persistent teammate treatment
 - own Workspace account, email (@agents.company.com), calendar, Drive
 - acts under its own identity, not yours → audit trail names the agent
 - unlike per-job temporary subagents, it keeps living and working
 - cloud-persistent: hour- and day-long jobs continue after you close the laptop

Four memory types:
 session (current job) · semantic (knowledge from docs and people)
 procedural (how work gets done, self-written skills) · episodic (job history)

Google also stresses full business context — pricing, portfolios, departmental norms grounded into answers. That goes beyond the internal-docs chatbot from retrieval into execution, continuing the classic-to-agentic RAG arc.

Skills and data ship as a set.

Skills: reusable work bundles (global library + company registry + self-made)
 → "turn this job into a skill" saves a reusable template
 → domain skills first for data and ML (PySpark code, notebooks, training, fixes)

Tools: enterprise registry (Salesforce, ServiceNow, Jira, Git, ...)
 → Confluence, Office, Teams, Slack, Workspace connections
 → BigQuery, Databricks, Postgres, Snowflake, desktop files, any MCP server

Three data services:
 Knowledge Catalog (map business definitions once) · Smart Storage (enrich unstructured)
 Borderless Lakehouse (query S3 and Azure Data Lake with no variable egress fees)

Watch the saved BigQuery reports for business users: build once, rerun without token costs, same verified answer every time. Recurring reporting off metered inference.

Cost and governance: what CFOs and security teams read

Agents hammering tools for hours scare budgets. Google offers three levers.

Lever Content
Multi-model orchestration small pieces on Flash-class models, hard parts on Argon-class (Gemini + Claude today, more planned)
Smart routing each workload auto-triaged to the cheapest capable model
Real-time spend caps per-project hard limits in Cloud Billing; agents pause when tripped (resume needs approval, chargeback by department) 
Governance in four sentences (who, may-do, trace, never):
 identity — cryptographically attested agent identity, least privilege
 authorization — role-based, mapped to external systems via OAuth
 auditing — every action logged to the agent, not a person
 policy — runs in an Agent Sandbox, all traffic via Agent Gateway (AI firewall)

Numbers shared: 1B monthly Gemini users, ~90% of the Fortune 100 on Gemini Enterprise, ~500 customers past one trillion tokens each. Early testers On, Shopify, PayPal; named customers BNP Paribas, Bradesco, Merck; cases like Commerzbank cutting document review from 20 hours to one and SOMPO fielding 10,000+ custom agents. Financial-services and legal specializations in preview; government, healthcare, retail next.

Three things are still missing: benchmarks, pricing, GA dates. Currently private preview, with industry expectations around end of October to early November. Consumption-based pricing with no extra SKU is all that is confirmed. A coworker agent holds a Workspace account with reportedly no per-seat charge — the metering unit will decide the real math.

CodeBridge Mini Lab: the pre-adoption checklist

Straight from the briefing: check permissions, scope, and money before the demo.

Data access:
 [ ] what the agent may read (cut by department, folder, database)
 [ ] a shared definition layer like a Knowledge Catalog
 [ ] sensitive data excluded or masked

Connector scope:
 [ ] which systems connect (Workspace, Slack, Jira, Salesforce, DBs…)
 [ ] each connector permissioned separately from human permissions
 [ ] agent-identity actions audited apart from humans

Cost and approval:
 [ ] per-project caps with departmental chargeback, pausing when tripped
 [ ] routing choice between automatic and manual
 [ ] non-metered options for recurring work like saved reports

Operations:
 [ ] coworker account provisioning and offboarding fits HR and IT
 [ ] mid-job reports and a cancel button for hour-long work
 [ ] industry skills for your vertical, as with finance and legal

Conclusion: one window, four things to judge

One line to close.

Judge agent platforms not by demos but by permissions, scope, money, and traces.

The Gemini agent message is clear: work starts in a prompt window, but the contest is decided behind it — connections, memory, routing, governance. Picked like a chatbot it fails; vetted like a hire it shows a path. Confirm data access, connector scope, spend caps, and approval flows. Then start the pilot.

Further reading

References

Go deeper with a course

To connect document search and answer generation to real work, this internal-docs chatbot course continues exactly where the adoption checklist here leaves off.