Gartner projects that by 2028, most AI agents will act by coordinating with other agents, not with people. Nobody has figured out how to govern a conversation you are not part of, and that is the next big exposure hiding inside the agent hype.
Gartner's new Agentic AI Hype Cycle drops a prediction that should make anyone responsible for risk put their coffee down.
By 2028, 60 percent of agents will autonomously interact with other agents.
Read that again slowly. The majority of what your AI agents do will soon happen in exchanges you never see, at machine speed, agent to agent, with no person in the loop for each handoff. Gartner has a name for where this goes: the agent-to-agent economy. Today your agent automates a task a human used to click through. Tomorrow it negotiates, delegates, and hands work to another agent, which hands it to another.
That is a genuine leap in capability. It is also a genuine leap in exposure, and the second part is getting almost none of the airtime.
Why a conversation between machines is a governance problem
When Agent A trusts Agent B's output and acts on it, and Agent B was wrong, spoofed, or out of date, the mistake propagates before a human ever sees it. There is no natural pause. The error just moves, fast, through a chain of automated decisions, and by the time it surfaces it has already touched three systems.
Three things break specifically, and Gartner keeps circling all three:
Identity. Which agent is this, really, and is it actually allowed to do what it is asking to do? In a world of agents calling agents, "trust me, I am an internal service" is not good enough.
Traceability. When something goes wrong across a five-agent handoff, can you reconstruct what actually happened, in order, with evidence? If the answer is no, you do not have a system, you have a rumor.
Accountability. When an autonomous chain produces a bad outcome, who owns it? "The AI did it" is not an answer a regulator, a client, or a board will accept.
Here is how the risk profile shifts the moment agents start coordinating:
| One agent, human-supervised | Agents coordinating, machine-speed |
| A person checks the output | The output triggers the next action automatically |
| One identity to trust | A chain of identities to verify |
| One log to read | A distributed trail across many agents |
| Clear owner | Diffused, unless you designed for it |
The takeaway
You do not govern this by writing a policy document after the agents are already talking. By then the conversation is happening thousands of times a day and your PDF is irrelevant. You govern it by building identity, logging, and guardrails into the architecture before the handoffs start, so that every agent has a verifiable identity, every action leaves a trail, and there is always a human-owned line of accountability behind the automation.
That is unglamorous, foundational work, and it is exactly the work most firms skip in the rush to say they have agents. It is also what Ayvede designs: the trust and oversight layer underneath autonomous AI, so the agent-to-agent economy runs on rails instead of hope. Adoption with less risk, before the risk compounds.
Advise. Vet. Deliver. More at ayvede.com.
One thing to keep: the scary part of agentic AI is not one agent going rogue. It is a hundred well-behaved agents confidently acting on each other's mistakes, faster than you can watch. Design for that now, while your agents still mostly talk to you.