Agent Development Kit and A2A: Google wants to network AI agents

In collaboration with 50 industry partners, Google Cloud has introduced a new protocol called Agent2Agent (A2A) for its AI platform Vertex AI, which is designed to enable agent-based communication. There is also a new Agent Development Kit (ADK) for developers and the so-called Agent Garden, which contains ready-made agents including the necessary connectors to make it easier for devs to try out the technology.

The ADK is currently intended for Python, but other languages could follow. According to Google Cloud, it should be possible to create an AI agent “in under 100 lines of code”. Among other things, it is possible to set in advance how agents draw conclusions (reasoning) and which systems they are allowed to interact with – including with specific guardrails from which they should not be able to escape. Leaks of sensitive data should also be prevented in this way.

Standard communication with automatic monitoring

Interactions with the agents are possible via text, audio, and video. The Vertex AI platform offers over 130 basic models (Foundation Models) up to Gemini 1.5 Pro. In total, there are more than 200 models available in the so-called Model Garden. In addition to Google itself, the suppliers include Mistral (from Europe), Meta (Llama), Anthropic (Claude) plus smaller providers such as Qodo or AI21. In addition to A2A, data can also be sent securely via the Model Context Protocol (MCP), according to Google.

Deployment is possible in both Vertex AI and Kubernetes – directly into productive operation if desired. Content filters for the output of the agents, defined output limits and explicitly prohibited subject areas are intended to ensure brand-compliant behavior in the corporate context – although there is always the question of the extent to which prompt hacking remains possible. As agents (can) assume the identity of their users, there is a separate identity management system with associated permissions. It should be possible to automatically monitor the behavior of individual agents in real time, but Google Cloud has not yet provided any further details.

Software developer kits for agents are not a new development. For example, OpenAI released the competing Agents SDK for its GPT models in March. It can (also) be used for open-source models. Amazon, on the other hand, has its Bedrock agents, which are being improved step by step and are also capable of orchestration. With A2A, Google is aiming to standardize communication between agents. The protocol is explicitly designed to be open, which also enables parallel operation with MCP (which originally comes from Anthropic). MCP is used to connect agents with different tools and the associated resources, while A2A is intended to organize inter-agent communication. A client agent and a remote agent are used for this. The client agent knows the user requirements, the remote agent carries them out. Tasks are distributed in the form of “conversation units” (e.g., for querying data from people) or “work units” (tool deployment and similar).

SAP, Atlassian, Salesforce and co. meet McKinsey, BCG and other consultants

According to Google Cloud, the partners that are to support A2A as soon as possible include Box, Intuit, Cohere, Atlassian, MongoDB, Salesforce, ServiceNow, PayPal, and SAP. Implementers on Google's vendor list also include the major consulting firms, including BCG, KPMG, McKinsey, PwC, Wipro, Accenture, Deloitte, and Capgemini. They hope that this will enable them to put agency process optimizations on track more quickly – whatever this means in practice for the end user. “We believe that the A2A framework will bring significant added value for customers whose AI agents will now be able to work with all existing enterprise applications,” says Google Cloud.

Collaborative AI agents require “universal interoperability” to realize their full potential. A2A uses established protocols such as SSE, JSON-RPC and HTTP. According to Google Cloud, authorization and authentication are on par with what OpenAI offers. A2A and the ADK are designed to enable “true multi-agent scenarios”. This means that agents are no longer limited “tools”. A2A can complete tasks quickly as well as long-running tasks, including in-depth research, “which takes several hours or even days”. However, a “human in the loop” is then mandatory. Real-time feedback is provided, including via a dedicated notification log. Google has not yet provided information on costs – i.e., how A2A and ADK will be integrated into Vertex AI operations –. A draft specification can be found on GitHub, along with sample code. Further details and a “production-ready” version of A2A are expected in the coming months – Google Cloud is relying on the aforementioned partners for implementation. “AI agents offer a unique opportunity to help people be more productive by performing many repetitive or complex daily tasks autonomously,” the company hopes.

(bsc)