#Agent Engine

[govindhtech]

Introduction to Multi Agent Systems Enhancement in Vertex AI

Multi-agent systems introduction

Vertex AI offers new multi-agent system creation and management methods.

All businesses will need multi-agent systems with AI agents working together, regardless of framework or vendor. Intelligent systems with memory, planning, and reasoning can act for you. They can multi-step plan and complete projects across many platforms with your instruction.

Multi-agent systems require models like Gemini 2.5 with better reasoning. They need corporate data and process integration. Vertex AI, its comprehensive platform for coordinating models, data, and agents, seamlessly integrates these components. It combines an open approach with strong platform capabilities to ensure agents work reliably without disconnected and brittle solutions.

Today, Google Cloud unveils Vertex AI advancements so you can:

Develop open agents and implement corporate controls

The open-source Agent Development Kit (ADK) is based on Google Agentspace and Google Customer Engagement Suite (CES) agents. Agent Garden has several extendable sample agents and good examples.

Vertex AI's Agent Engine is a managed runtime that safely deploys your custom agents to production globally with integrated testing, release, and reliability.

Connect agents throughout your organisation ecosystem

The Agent2Agent protocol gives agents a single, open language to communicate regardless of framework or vendor. This open project is led by us and collaborates with over fifty industry professionals to further our multi-agent system vision.

Give agents your data using open standards like Model Context Protocol (MCP) or Google Cloud APIs and connections. Google Maps, your preferred data sources, or Google Search may power AI responses.

Creation of agents using an open methodology with Agent Garden and Agent Development Kit

Google's new open-source Agent Development Kit (ADK) simplifies agent creation and complicated multi-agent systems while maintaining fine-grained control over agent behaviour. You can construct an AI agent using ADK in under 100 lines of user-friendly code. Look at these examples.

Available currently in Python (other languages will be released later this year), you can:

With orchestration controls and deterministic guardrails, you can accurately govern agent behaviour and decision-making.

ADK's bidirectional audio and video streaming enable human-like agent conversations. Writing a few lines of code to establish genuine interactions with agents may turn text into rich, dynamic discourse.

Agent Garden, a suite of useful tools and samples in ADK, can assist you start developing. Use pre-built agent components and patterns to learn from working examples and expedite development.

Pick the model that fits you. ADK works with all Model Garden models, including Gemini. Anthropic, Meta, Mistral AI, AI21 Labs, CAMB.AI, Qodo, and others provide over 200 models in addition to Google's.

Choose a deployment destination for local debugging or containerised production like Cloud Run, Kubernetes, or Vertex AI. ADK also supports MCP for secure data-agent connections.

Launch production using Vertex AI's direct integration. The reliable, clear path from development to enterprise-grade deployment eliminates the difficulty of transitioning agents to production.

ADK is optimised for Gemini and Vertex AI but works with your chosen tools. Gemini 2.5 Pro Experimental's improved reasoning and tool-use capabilities allow ADK-developed AI agents to decompose complex challenges and communicate with your favourite platforms. ADK's direct connection to Vertex AI lets you deploy this agent to a fully controlled runtime and execute it at enterprise scale.

Agent Engine deploys AI agents with enterprise-grade controls

Agent Engine, Google Cloud's controlled runtime, simplifies AI agent building. Agent system rebuilding during prototype-to-production is no longer required. Agent Engine manages security, evaluation, monitoring, scaling complexity, infrastructure, and agent context. Agent Engine integrates with ADK (or your chosen framework) for a smooth develop-to-deploy process. Together, you can:

Use ADK, LangGraph, Crew.ai, or others to deploy agents. Choose any model, such Gemini, Claude from Anthropic, Mistral AI, or others. Flexibility is paired with enterprise-grade control and compliance.

Keep session context: The Agent Engine supports short-term and long-term memory, so you don't have to start over. This lets your agents remember your discussions and preferences as you handle sessions.

Vertex AI has several tools to evaluate and improve agent quality. Improve agent performance by fine-tuning models based on real-world usage or utilising the Example Store.

Linking to Agentspace can boost utilisation. You may register Agent Engine-hosted agents with Google Agentspace. Gemini, Google-quality search, and strong agents are available to employees on this corporate platform, which centralises management and security.

Google Cloud will improve Agent Engine in the next months with cutting-edge testing and tooling. Agents can utilise computers and programs. To ensure production reliability, test agents with many user personas and realistic tools in a specialist simulation environment.

The Agent2Agent protocol connects agents across your enterprise

One of the biggest barriers to corporate AI adoption is getting agents from different frameworks and suppliers to work together. Google Cloud worked with industry leaders that support multi-agent systems to create an open Agent2Agent (A2A) protocol.

Agent2Agent allows agents from different ecosystems to interact, regardless of framework (ADK, LangGraph, Crew.ai, etc.) or vendor. A2A lets agents securely cooperate while publicising their capabilities and choosing how to connect with users (text, forms, bidirectional audio/video).

Your agents must collaborate and access your enterprise truth, the informational environment you developed utilising data sources, APIs, and business capabilities. Instead of beginning from scratch, you may give agents your corporate truth data using any method:

ADK supports Model Context Protocol (MCP), so your agents may connect to the growing ecosystem of MCP-compatible devices to access your numerous data sources and capabilities.

ADK lets you directly connect agents to corporate capabilities and systems. Data from AlloyDB, BigQuery, NetApp, and other systems, as well as more than 100 pre-built interfaces and processes established using Application Integration, are included. Your NetApp data may be used to create AI agents without data duplication.

ADK makes it easy to connect to call tools from MCP, LangChain, CrewAI, Application Integration, OpenAPI endpoints, and your present agents in other frameworks like LangGraph.

We manage over 800K APIs that operate your organisation within and outside Google Cloud. Your agents may utilise ADK to access these API investments from anywhere with the correct permission.

After linking, you may supplement your AI replies using Google Search or Zoominfo, S&P Global, HGInsights, Cotality, and Dun & Bradstreet data. For geospatial agents, we now allow Google Maps grounding. To maintain accuracy, we refresh 100 million Maps data points daily. Grounding with Google Maps lets your agents reply with geographical data from millions of US locales.

Create trustworthy AI agents with enterprise-grade security

Incorrect content creation, unauthorised data access, and prompt injection attacks threaten corporate AI agents' functionality and security. Google Cloud's Gemini and Vertex AI building addresses these difficulties on several levels. You could:

Manage agent output with Gemini's system instructions that limit banned subjects and match your brand voice and configurable content filters.

Identity controls can prevent privilege escalation and inappropriate access by determining whether agents work with dedicated service accounts or for individual users.

Google Cloud's VPC service controls can restrict agent activity inside secure perimeters, block data exfiltration, and decrease the impact radius to protect sensitive data.

Set boundaries around your agents to regulate interactions at every level, from parameter verification before tool execution to input screening before models. Defensive boundaries can limit database queries to certain tables or use lightweight models with safety validators.

Automatically track agent activities with rich tracing features. These traits reveal an agent's execution routes, tool choices, and reasoning.

Build multi-agent systems

Vertex AI's value depends in its whole functionality, not simply its features. Integration of solutions from various sources is now easy on a single platform. This unified strategy eliminates painful model trade-offs, corporate app and data integration, and production readiness.

[deep4ried]

Who says machines don’t have souls? / Aren’t we machines, too?

[heph]

D3: Alternate Universe - SONIC BOOM

[badolmen]

They don’t even attempt to assassinate US politicians anymore. You notice that? Not since the anthrax scare back for… who was it, Barack? And even that… pathetic. This new generation has no respect for an honest hitman. I’m not sure this new generation has any honest hitman - you see that shit with Boeing? Sloppy, fucking disgraceful - you kill the whistleblowers before they get halfway to a lawsuit. What kind of fucking amateur is doing faked suicides the night before testimony? Goddamn greenhorns. Back in my day someone tried to shoot Ronald Reagan in broad daylight. There used to be bomb threats to Congress. I took out a few union leaders in the utilities sector myself. Today’s generation? Won’t even threaten to throw a punch - not even over on that - what’s it now, ‘X’? They got no guts. None! And they don’t even have poor impulse control to boot! Too much of that - that panopticon anxiety bullshit. “Oh what if I get a called out post???” People used to send the president letters full of bioweapons. In the mail! Today’s generation? Not a chance. All because of woke.

[bearloonz]

Everything will always return 2 them

[dam-thats-blue]

Might as well post here too since Twitter is loving it, pick your poison.

[sleepwalk-station]

failure

[myxineye]

pixel tests :p

[gophergal]

there are many kinds of love

TFTober 3 - Relationships

[fluffydeoxys]

a joke concept that i got attached to, whether it'll stick or not remains to be seen. but there are tiny bits of lore for them that seeeeems potentially interesting. we'll see

[sugarrcola-blog]

dei and doc, all lazy sketches

[vommitdurilka]

GOOD MORNING GAY NATION!!!!!!!🏳️‍🌈🏳️‍🌈🏳️‍🌈

[catsurfavs]

Headcanon....

[nevadas-explorer]

i post again after 5000 years...

color refs for various flavors of guy

[selemina]

I finally watched Animator VS Animation. I uh... I have feels. Those are some EXPRESSIVE stickmen, damn! If anything, AVA is a masterclass on body language! But it also happens to be fun, spectacular, emotional and beautiful! So much effort goes into every episode, I as a fellow animator I am measuring the amount of work and I am HORRIFIED! So it is time for another episode of "Sel finds stuff on the internet" : Go watch AVA! You won't be disappointed!

[kakmem]

random doodles ft. 2x spy and engie

+ close-ups