A：The Gemini Enterprise agent platform is a tool that allows developers to build autonomous agents that can proactively assist users and complete tasks independently. It is powered by Gemini models, including Pro and Flash models, and supports the integration of other models like Clad from the model garden. The platform provides capabilities for building, scaling, and optimizing agents, including an agent development kit (Go) for skills and tools, communication with Google Cloud services, scalable serverless runtime, unique agent identity, policies for connecting and collaborating with other agents, and agent observability for monitoring and management.

A：The agent platform supports the construction and operation of agents through several features: the agent development kit (Go) to build agents with necessary skills and tools; scalable serverless runtime to keep agents connected and personalize context with sessions; unique agent identity and policies for secure governance; an agent registry and collaboration protocol for agents to discover and work with each other; and observability for managing, monitoring, and optimizing agents based on production activity.

A：The shared context across Gemini Enterprise Workspace and third-party agents is significant as it allows the entire ecosystem of agents to collaborate with each other using a common shared context. This context is provided by Google and can be used in both Gemini Enterprise Workspace and third-party agents, enhancing coordination and efficiency in their operations.

A：The marathon planning example used to demonstrate the capabilities of the agent platform involves planning a marathon in Las Vegas, simulating its impact on the city, and assessing different scenarios to determine the best course of action. This involves creating simulations to understand potential impacts such as rerouting emergency services, boosting the local economy, and showcasing iconic landmarks. Three main agents are used: the planner to determine race routes, the evaluator to assess routes based on business and community requirements, and the simulator to take the route and create actors with randomized behaviors to see the net impact to the city.

A：In the marathon simulation, the role of the planner agent is to determine the optimal routes for the marathon. It uses skills for handling geospatial data, mapping, and calculating viable routes to plan the path that the race will take. The planner agent's instructions and capabilities are designed to understand its role as the marathon planner and to coordinate with other agents involved in the simulation, such as those managing hydration stations, medical tents, and traffic flow.

A：The simulation app showcasing the marathon planning is powered by different technologies depending on the interface: Angular for the base and 3D animation, and Flutter for the sidebar and UI interactions. The app uses various Google products including Gemini, Firebase, and AI to render the simulation and provide a realistic race environment.

A：The agents were given instructions on marathon planning that were crafted with Gemini's help.

A：The plan is grounded in real data by making the agent an expert with tools and skills using maps and races, achieved by adding tools and skills, and downloading pre-populated Python code designed by the agent designer.

A：The agent is given access to mapping tools using Google Cloud MCP server for Google Maps and is provided with skills to answer map-related questions and find viable routes with starting and ending locations.

A：The mapping skill allows the agent to selectively call Maps tools based on the skill being loaded, while the GIS skill handles geographic information and includes access to Python scripts for processing Geojson data and finding viable routes.

A：The race director skill involves using a Google Doc created by the race planning committee, which was converted to a scale using Gemini. This skill allows agents to plan and check marathon routes based on historical requirements.

A：It takes 4 to 5 minutes to deploy the agent, which includes the skills necessary to create route plans for marathon participants.

A：The purpose of the team of agents is to distribute different responsibilities based on use case, access required, and capabilities needed, just like a team working together.

A：The two components that facilitate agents to communicate with each other are the A2A (agent to agent) protocol, which is a universal protocol for agents to advertise their capabilities and communicate, and the agent registry, which serves as a central directory resolving every agent's identity and their skill set.

A：The dynamic UI in the system is built by the agents using an open standard created by Google, which allows them to create dynamic, expressive UI that moves beyond walls of text. This enables the agents to design and build the exact interface they need using a common design language.

A：The simulation agent leads the process of running actual plans, working with the planner to get approved routes, and showing results. It involves connecting the simulator with the planner, and uses the A2A protocol and agent registry for coordination.

A：The significance of the agent registry and A2A protocol in the agents' collaboration is to eliminate brittle API code and provide a system where agents can read each other's capabilities. The registry serves as the DNS of the internet of agents, resolving every agent's identity and mapping their specific skill set across the entire network, while the A2A protocol allows agents to connect and communicate without the need for custom coding or dashboards.

A：Agents have been utilizing real-time evaluation, dynamic UI for self-rendering, automatic connections using 8-way, and the agent registry to improve user experience.

A：In the past, the approach to optimize agents for the next run involved stashing the complete history of prior runs into each new request. Now, the process has evolved to using architectural approaches and database tools for more efficient and thoughtful management of data.

A：Stateful agents and context windows are significant in agent engineering as they enable agents to become stateful over time and modify their context as needed, improving their ability to manage information available to them and their own generalized knowledge to complete tasks.

A：The memory bank contributes to an agent's performance by enabling it to pull learnings from prior events into the context window relevant for specific tasks and store memories for future sessions. It also allows the use of tools like Spark or alloy DB to provide additional context to agents using a differentiated data approach.

A：The planner agent plays a role in demonstrating context and memory improvement by managing sessions, adding memory, and accessing other data sources. By integrating with agent platform sessions and memory bank, the planner agent can remember details from previous plans and recall them for future planning.

A：Unstructured data, such as local rules, is utilized by the planner agent through data engineering processes, including chunking the documents, transforming them into embeddings, and using retrieval augmented generation (RAG) to query and integrate this data into the planning process.

A：The purpose of managing context and adding memory to the planner agent is to enable the simulation to adjust its route based on the latest context and data, and to allow the planner agent to recall details from its long-term memory of past simulations, improving the intelligence and adaptability of the system.

A：The speaker illustrates the benefit of context engineering by demonstrating how a simulation adjusted its route based on new data and context, which reflects the intelligence gained from managing context and adding memory to the planner agent.

A：To optimize and debug agent behavior, tools and techniques being used include agent observability for visibility into operational metrics, Gemini Cloud Assist for managing agents from design to optimization, and the Cloud Monitoring console to identify and investigate issues such as high latency or tool calls that lead to system failures.

A：The simulator agent failed to call the Gemini Model API due to a request error, which indicated that something was wrong with the payloads sent to the model.

A：Cloud Assist pointed to the specific line of the agent simulator code where the issue occurred, allowed the user to continue the investigation started in the console, and provided explanations and suggestions for resolving the 400 error caused by exceeding the Gemini API's context token limit.

A：Agents not only find the root cause of issues but also suggest proactive code fixes. In this case, the agent proposed a fix by suggesting an adjustment to the event compaction configuration.

A：The debugging experience with Gemini Cloud Assist and the agent helped in tracing the root cause of the failure, identifying the issue with the storage not loading the model fast enough, and proactively fixing it with the help of Cloud Assist.

A：The architecture for applications and AI systems evolves based on learning from AI's performance and needs. It may involve changes like swapping out runtimes, databases, or programming languages as the application changes.

A：Gemini Cloud Assist helped upgrade an existing system in the simulation by converting a Cloud Run service to a Google Kubernetes Engine (GKE) deployment and introducing a fine-tuned or customized model for the same cluster.

A：The envisioned relationship between agents and the implementation of change in the autonomous cloud is one where agents work with humans not just to answer questions but also to partner with them in safely implementing changes.

A：Users can create their own no-code agents in Gemini Enterprise by utilizing the agent runtime, which allows them to make work easier, including tasks like logistics, ordering supplies, and managing various aspects of events.

A：Agents in Gemini Enterprise have the capability to plan and order for races, work together to create comprehensive plans, and manage logistics including water, food, and entertainment. They also have access to dynamic interfaces and can be given documents as context to aid in their functionality.

A：The role of the agent registry in Gemini Enterprise is to keep a record of all agents that have been created, making them discoverable by other agents and apps included in the Gemini Enterprise platform.

A：The marathon planner agent in Gemini Enterprise is used to facilitate the planning and ordering of events, such as marathons, by automating tasks like ordering supplies, managing logistics, and ensuring necessary items like water and portable toilets are available.

A：Gemini Enterprise handles security concerns by implementing agents that provide users and other agents new ways to intentionally or unintentionally expose data. It features agent identity and policies to ensure agents can only access what they need when they need it, and an agent gateway for managing access control policies. The platform also includes the ability to set conditions on agents' access to resources like financial databases, and a Zero Trust architecture that enforces role-based access control using agent identities and policies.

A：The tools presented, such as the Wiz platform, can scan code and infrastructure to build a security graph, identify risks and vulnerabilities, and suggest remediations directly within developer tools. This allows developers to ensure their code is secure and address potential threats early in the development process.