'A collection of safe agents does not make a safe collection of agents'
A new government-commissioned report has flagged the growing risk of deploying multi-agent AI systems in the wake of their rising adoption in workplaces.
An AI agent is a system that interacts with an environment through observation and actions, with the objective of achieving a specific goal, according to the report from the Gradient Institute.
It is distinguished by its AI-driven adaptive decision-making engine, rather than following a set plan.
"This means that AI agents are able to handle greater variability, but often at the cost of reduced reliability and interpretability," read the report, commissioned by the Department of Industry, Science, and Resources.
In Australia, a SnapLogic poll found that 74% of business leaders consider the implementation and development of AI agents over the next 12 months as a priority.
Deployment of multi-agent systems
But the Gradient Institute's report found that single-agent deployments are evolving into multi-agent deployments that can interact with each other.
It cited as an example employers' use of HR agents to process new hires, which must coordinate with IT agents for provisioning access and finance agents for setting up payroll.
"Such cross-functional dependencies suggest that enabling direct agent-to-agent communication could streamline operations beyond what isolated agents could achieve," the report read.
But Dr Tiberio Caetano, Gradient Institute's Chief Scientist and co-author of the report, pointed out that the use of multi-agent systems is a "fundamental shift in how organisations need to approach AI risk and governance."
"A collection of safe agents does not make a safe collection of agents," Caetano said in a statement. "As multi-agent systems become more prevalent, the need for risk analysis methods that account for agent interactions will only grow."
Risks of agents
Among the risks identified in the report are:
- Inconsistent performance of a single agent derailing complex multi-step processes
- Cascading communication breakdowns as agents misstate or misinterpret messages
- Shared blind spots, and repeated mistakes, when a team of agents all use similar AI models
- Groupthink dynamics where agents reinforce, rather than critique, each other's errors
- Coordination failures when agents don't understand what their peers know or need
- Competing agents optimising for individual goals that undermine organisational outcomes
The report offers a toolkit for organisations to identify and assess risks for multi-AI agent systems.
It also recommended progressive testing stages for multi-agent systems to increase exposure to potential impacts.
"By providing practical tools grounded in rigorous science, we're enabling organisations to better understand the novel risks that emerge when AI agents work together - and how to start addressing them," said Bill Simpson-Young, CEO of Gradient Institute, in a statement.
"The path forward isn't about avoiding this technology; it's about deploying it responsibly with awareness of both its potential and its pitfalls."
