AI that speaks, codes and takes action: Inside Google Cloud’s Gemini masterclass

Google Cloud demonstrated Gemini acting as a worker, not only a conversational answer engine. It could analyse large collections of information, create presentations, generate software, check code, and produce a detailed 3D scene. This matters because businesses could delegate complete tasks instead of requesting one answer at a time.

One demonstration fed Lumen,000 files from NASA’s open archives. The system used Mars weather readings and helicopter telemetry to create a presentation with a supplied template and design guidelines. Another generated a 3D library scene with Python and NumPy, without dedicated rendering software or outside assets. CodeMender also identified vulnerabilities, created patches, and tested them in a sandbox.

These examples show AI handling research, design, coding, and security work. The demonstrations also point to a practical limit: companies must provide suitable knowledge, tools, permissions, and safeguards. Shivani stressed that simpler tasks can use faster, cheaper models rather than a frontier model every time.

Agentic AI is a system that can pursue a goal through several steps. It can plan, find relevant information, use software tools, and take action for a person or organisation. A chatbot generally responds to prompts, while an agent can continue working toward an outcome. This makes AI useful for processes, not just conversations.

Rajesh Shivani described an agent as “a model plus a harness.” The model provides intelligence, while the harness manages memory, context, permissions, and task execution. Useful agents also need a clear goal, instructions, relevant knowledge, and tools. Company data matters because general-purpose models may not know internal processes or customer information.

The shift changes how businesses organise work. Instead of asking AI for isolated answers, they can delegate tasks such as preparing reports, coordinating interactions, or submitting a banking request. However, agents must operate reliably and safely. Organisations need controls for access, security, costs, and automated activity as agents begin interacting with business systems.

The demonstration showed that these systems can work across very large collections of information, but the article gives one specific scale rather than a universal maximum. Gemini analysed 80,000 files from NASA’s open archives. The important point is not only the file count. It is that the system turned scattered material into a useful finished product.

The files included Mars weather readings and helicopter telemetry. Lumencessed that information and produced a presentation with charts, images, and a clear narrative, using a supplied template and design guidelines. This shows how an agent can combine information retrieval, analysis, writing, and presentation design in one workflow.

For businesses, the example suggests a way to convert large internal archives into reports or other outputs. Results still depend on the knowledge, instructions, and tools given to the agent. The demonstration also supports Shivani’s broader message: AI can take on complex work, while companies must manage permissions, security, costs, and reliability.

Voice AI can handle practical conversations, translation, and service tasks. It is moving beyond simple speech responses toward helping people complete real-world interactions. Shivani described voice as becoming a new user interface, because customers and employees can use spoken communication to access services and information.

A Grab system helped passengers and drivers coordinate airport pickups despite language barriers, with support for more than 85 languages. Another demonstration dubbed an English video into Marathi and Malayalam while matching lip movements and facial expressions. A Citibank virtual banking assistant explained choices for a maturing certificate of deposit and submitted a renewal after the customer selected a 12-month term.

These examples cover travel coordination, multilingual media, and banking service. They show voice AI handling both communication and action. Businesses could use similar capabilities for training, marketing, customer support, and employee interactions. The article presents these as demonstrations of a broader shift toward voice-based interfaces.

Giving agents internal data, software tools, and permission to act changes AI from an answer system into a work system. The agent can use company knowledge to understand a process, choose steps, and complete an outcome. This matters because general-purpose models may not know an organisation’s specific procedures or customer information.

The demonstrations show what this can look like. Gemini turned NASA’s archive into a presentation, while the Citibank assistant explained certificate-of-deposit options and submitted a renewal after the customer chose a term. Agents need a goal, clear instructions, relevant knowledge, and tools. A harness manages memory, context, permissions, and task execution around the model.

The benefit is delegated work, but access creates new responsibilities. Businesses must rethink permissions, security, costs, and automated activity. They also need reliable safeguards because agents, rather than only people, may interact with company systems. Developers could increasingly manage agents instead of performing every individual step themselves.

The most powerful AI model is not automatically the best choice for every business task. Many jobs are simpler and do not require maximum capability. Using a smaller model can help an organisation control costs and get faster responses. This makes AI adoption more practical across everyday workflows.

Rajesh Shivani specifically advised that companies do not need a frontier model for everything. A business might reserve a powerful model for difficult analysis, while using a faster, lower-cost model for simpler tasks. The article does not list a detailed model-selection process, but it clearly links simpler tasks with cheaper, faster models.

This approach treats model choice as part of business design rather than a contest for the biggest system. Companies can match capability to the work being done. They still need to consider reliability, permissions, security, and tools, because a less expensive model does not remove the need for safe and well-managed agent workflows.

An AI agent needs more than a model that generates text. The model supplies reasoning and language ability, while the surrounding harness organises the work. It manages what the agent remembers, the context it can use, the permissions it has, and how tasks are executed. This structure lets AI move from answering to acting.

The agent begins with a goal and clear instructions. It then uses relevant knowledge, such as a company’s internal information, and tools that let it perform actions. For example, a banking assistant could explain certificate-of-deposit choices and submit a renewal after the customer selects a term. The harness controls the information and permissions involved.

This combination makes agents useful for multi-step jobs, but it also creates operational risks. Businesses must ensure that agents access the right information and tools, complete tasks reliably, and act safely. They need to plan for security, costs, permissions, and systems increasingly interacting with agents rather than only people.