A2A Protocol
Agent2Agent (A2A)
What is the A2A protocol?
- A shared standard for agent collaboration
- Complements MCP (agent↔tool) with agent↔agent communication
- The foundation for automation that combines agents from multiple vendors
What A2A really changes is the unit of integration work. Until now, connecting another vendor’s agent for a department meant getting the other side’s API documentation and building a one-to-one connector, and with five agents you had to manage twenty connections separately. With A2A, each agent publishes a spec in which it declares what it can do, and the other side simply reads that spec to hand off tasks. The difference you feel is integration moving down from a development project to a registration task.
In practice you will meet three main components. The agent card is a business card listing the address, capabilities, and authentication method, posted as a JSON file at a fixed path on the server. A task is a unit of work that moves through states such as received, in progress, input required, and completed, so everything from a 3-second inventory lookup to a three-day review is handled in the same format. Results come back split into messages for conversation and artifacts for deliverables. If MCP is the socket where you plug in tools, A2A is the approval chain along which colleagues hand work to one another, so having a single agent use both standards at once is the default setup.
The prerequisites are clear too. A standard only works if the other side adopts it, so systems that don't speak A2A, such as an in-house RPA bot or a 20-year-old ERP, need an adapter layer. Before drawing the collaboration structure, decide which agent makes the final call. Connect equal peers with no coordinator and you get automation with no one holding accountability.
The most common misconception is that attaching a protocol makes agents coordinate on their own. In practice, the first thing that goes wrong is a delegation loop, with agent A asking B and B asking A back, leaking hundreds of thousands of won in API costs overnight. Without delegation depth limits, per-request call and cost caps, and end-to-end audit logs, you can't trace which agent produced a number even after a wrong quote reaches a customer. Where customer data passes to an external vendor's agent, specify the scope of data processing in the contract before you start.
John KimInstructor John Kim, author of B2B Marketing Design and CEO of Nanum Consulting
Seolyeong ParkGenerative AI and AX training specialist with 17 years of hands-on experience and 200+ lectures, helping non-technical staff put AI to work right awayOnce MCP, which connects agents to tools, became widespread, attention turned to standardizing how agents work with each other. In 2025, Google and others released the A2A (Agent2Agent) protocol.
Let’s look at a hypothetical case. Company A, an industrial materials distributor, has 340 employees and annual revenue of KRW 180 billion (≈ US$134,670,100). It receives 620 client quote requests a month; four sales-support staff were swamped and still took an average of 2.5 days to reply. The sales team used an external SaaS quoting agent and the logistics team used an in-house inventory-lookup agent, and a custom development quote to link the two came to KRW 60 million (≈ US$44,900).
- Manager Kim on the sales support team said, “I message the inventory team three times just to produce one quote.” The team broke down the wait times in 8 weeks of processing logs (620 requests) and found that 63% of the delay came from waiting on inventory and delivery-date checks.
- Instead of rebuilding the two agents, Park, the IT Planning team lead, changed course to attach an A2A agent card to each one and publish its capabilities and authentication method, and the development quote dropped from KRW 60 million (≈ US$44,900) to KRW 14 million (≈ US$10,500).
- In a two-week pilot, the quote agent passed the inventory agent a task: “Part KX-220, 300 units, can you deliver within 3 weeks?” The inventory agent sent it back with an “input required” status because the shipping origin was missing.
- In the first week, the two agents fell into a loop of re-confirming with each other and daily API costs hit KRW 410,000 (≈ US$305). The team capped delegation depth at 2 levels and calls at 8 per request, and any request over the limit now goes to Manager Kim.
- The legacy ERP couldn’t speak A2A, so an adapter for inventory master lookups was layered on top, and Lee, a deputy manager on the logistics team, wrote the rule "agents never finalize unit prices" into the card description.
- After three months, the average quote reply time fell from 2.5 days to 4 hours, 71% of 620 monthly requests closed with a single manager check, and inventory-inquiry messages dropped from 480 to 90 per month.