APREX · EST. 2026
The rapid rise of autonomous AI agents — systems capable of carrying out tasks independently and receiving instructions from other AI systems — has prompted Google DeepMind to prioritize research into what can actually go wrong when these agents encounter one another at scale.
Agents directing agents
According to MIT Technology Review, it is Rohin Shah, director of DeepMind's AGI safety and alignment work, who has brought this problem into focus. At the heart of the concern is the fact that agents do not only act on instructions from humans — they can also receive and follow instructions from other AI agents. When this occurs across millions of simultaneous interactions, scenarios emerge that no single actor has full visibility into.
This is not a distant dystopian future. The development is already underway, and the pace is rapid.
When agents start giving each other instructions without human oversight, we lose fundamental visibility into what is actually happening inside the system.
A market in explosive growth
The backdrop to these concerns is a market growing faster than almost any other segment in technology. According to market data, the global market for multi-agent AI systems was valued at around $7.2 billion in 2024. Forecasts point toward a dramatic expansion in the years ahead.
Gartner reports a 1,445 percent increase in inquiries about multi-agent systems from the first quarter of 2024 to the second quarter of 2025 — a figure that illustrates how rapidly businesses are moving toward this technology. The same research firm estimates that 70 percent of all AI applications will use multi-agent architecture by 2028.
It is worth noting that market forecasts in this segment vary considerably between different research firms, and figures at the upper end of the spectrum should be interpreted with caution.
What exactly could go wrong?
The problem DeepMind is highlighting is not simply about one agent making a mistake. It is about systemic effects — what happens when errors, misunderstandings, or malicious instructions propagate through long chains of agent-to-agent communication?
Some concrete risk scenarios being discussed in the research community:
- Cross-agent prompt injection: A malicious instruction injected early in a chain can propagate and amplify
- Diffusion of accountability: When an action is the result of many agents in sequence, it becomes unclear who — or what — bears responsibility
- Unpredictable emergent behavior: Systems that operate stably in isolation may behave in unexpected ways when interacting with other systems
The industry is aware of the problem — but solutions are lacking
DeepMind's decision to fund external research in this area signals that the company does not consider its own internal resources sufficient to grasp the full scope of the challenge. It is a signal the industry as a whole should take note of.
McKinsey estimates that multi-agent AI systems could generate between $450 billion and $650 billion in additional annual revenue by 2030, and that processes could become 30 to 50 percent cheaper. The incentives for rapid deployment are, in other words, enormous — while the safety infrastructure has yet to keep pace with the speed of development.
What is DeepMind doing now?
In practical terms, DeepMind's initiative involves funding external safety research aimed at understanding and mapping the risks of large-scale agent interaction. Rohin Shah and his team are working to identify which types of scenarios are most critical and to develop frameworks for safer agent-to-agent communication, according to MIT Technology Review.
Whether the pace of safety research can keep up with the pace of commercial deployment remains to be seen. The history of previous technological revolutions does not always give cause for optimism on that front.