TORONTO — Major tech firms are joining forces on a new system that lets their AI agents communicate with each other and work together, a critical step as automated assistants enter the workplace. 

Google on Wednesday announced the Agent2Agent (A2A) protocol, with backing from over 50 tech companies including Cohere, PayPal, Salesforce and Workday. “Customers are asking for these new agentic systems to work with each other,” said Joe Davis, executive vice-president of platform engineering and AI at ServiceNow, another A2A participant.

Agents using the protocol can give or accept tasks based on digital cards that list what each one can do. The assistants keep in contact with each other until an assignment is done, and can also access information about jobs that have been completed. “Customers can give their agent a task and it will automatically find and connect to everything—data, APIs and other agents—needed to do that task,” said Amin Vahdat, Google vice-president of machine learning, systems and cloud AI.

Davis cited the example of a worker experiencing an error while using a Google product, and tasking an AI agent to fix it. Google’s AI agent could figure out what the problem was, then work with ServiceNow’s AI agent to find the right patch as well as a scheduled maintenance window in which to run it. “Working across different systems can be automated 24/7 to drive down resolution times for customers,” he said.

Most major tech firms have built agents in their software platforms, letting users delegate tasks like doing research, fixing IT outages or sending sales messages. Typically built on top of large language models (LLMs), these assistants are only as useful as the data and systems to which they have access. That gives the A2A participants an incentive to let their agents work together, even though they may sell competing products; Google, Salesforce and ServiceNow all have automated tools for customer service, for example.

As AI agents become a core part of all software systems, “interoperability is critical,” said Autumn Moulder, vice-president of engineering at Cohere, noting the sector is “in a rapid period of expansion where multiple industry standards are evolving.” 

The Toronto-based firm’s recently launched North platform lets users build AI agents powered by its LLMs. The assistants perform tasks using information from clients’ databases and from their other software systems, which are connected via application programming interfaces (APIs). 

Moulder and Davis both said the rules for how agents work together and with other technology tools are still in their infancy. Protocols like A2A could become more useful as more firms buy in, since that lets the agents do more. But the system’s design means it “can deliver immediate utility, even as the network grows,” said Moulder.

Many tech firms are also participating in a different system created by Anthropic called the Model Context Protocol (MCP), which makes it easier for agents to access data from app and site APIs. Cohere, Google and ServiceNow are all using it, as are Amazon and OpenAI. The two protocols together “ensure AI agents have the right context and can leverage the most useful tools,” said Moulder.