In collaboration with 50 partners from the industry, Google Cloud has a new protocol which is named Agent2agent (A2A) for its AI platform Vertex ai It has been introduced that agents should enable communication. There is also a new agent for developers Development Kit (ADK) and SO -Called Agent Garden, already prefabricated agents, including essential connectors, making it easier to try the gods.
ADK is currently intended to python, other languages may follow. According to Google Cloud, it should be possible to create AI agent “in less than 100 lines of code. In advance it is adjustable, among other things, agents attract conclusions (arguments) and with which systems they can interact with – as well as with specific railings in which they should not extend. Leak sensitive data should also be prevented.
Communication according to standard with automatic monitoring
The interaction with agents is possible by reading lessons -both audio and video. As part of the vertex AI platform, Gemini is more than 130 basic models (foundation models). Overall, more than 200 models are available in the SO -collapsed model garden, next to Google itself is Mistral (from Europe), Meta (Lama), Anthropic (Cloud) Plus small providers such as cudo or AI21. In addition to A2A, according to Google, data can be sent safely via model reference protocol (MCP).
A deployment is possible in both vertex AI and Kuberanets – for productive operation at direct request. For brand-friendly behavior in the context of the company, material filters for output of agents, defined output boundaries, and clearly prohibited subject areas should be served-the question is always present to what extent the prompt hacking is possible. Since agents handle their users, they have their own identity management with the permissions concerned. The behavior of individual agents should be automatically monitored in real time, and Google Cloud has not yet provided more details.
Software developer kit is not a new development for agents. So there was competition in Openaai Agent for your GPT model SDK Published in March. It can be used for open source models. Amazon is his in return FastThose who step step by step and also master orchestration. With A2A, Google aims to standardize communication between agents. The protocol is clearly opened, which also enables parallel operations with the MCP (which originally comes from anthropic). MCP is used to combine agents with various devices and related resources, while A2A is considered to organize inter -communication communication. A customer agent and a remote agent are used. The client agent knows that the user needs, the distance agent takes it out. The functions are distributed as “conversation units” (eg to queries data in humans) or “work units” (using equipment and so).
SAP, Atlassian, Salesforce and Co. Meking mckinsey, BCG and other consultants are available
According to Google Cloud, partners who are going to support A2A as soon as possible include box, intuit, cohere, atlassian, mongodb, salesforce, servicenow, paypal and SAP. The vendors of Google are also very large advisory houses to apply who are in the list, including BCG, KPMG, McKinsey, PWC, Wipro, Accentne, Deloitte and Capgemini. There is a hope that the adaptation of the aging process on the track rapidly means in practice for the end users. “We believe that the A2A framework will bring an important additional value for customers whose AI agents will now be able to work in all existing corporate applications,” said Google Cloud.
The colleague AI agents required “Universal Interoperability” to be able to take advantage of their full potential. The A2A uses installed protocols such as SSE, JSON-RPC and HTTP. According to Google Cloud, there are authority and authentication at the level of providing Openai. A2A and Adk should enable “real multi-agent scenarios”. This means that the agents are no longer “equipment”. A2A can quickly complete complete tasks, as well as involve long -lasting tasks, including intensive research, “that needs several hours or even day”. A “human in a loop” is then mandatory. Real -time response is planned, including your own notification protocol. Information about costs as A2A and Adk is integrated in vertex AI operation-now not Google is not created Draft specification Can also be found on github, as well as sample codeFurther details and the “production maturity” version of A2A can be expected in the coming months – Google Cloud depends on the partners mentioned for implementation. “AI agents provide a unique way to help people make more producers by doing many recurring or complex tasks every day,” the company is expecting.
(BSC)
