Every week, the world is treated to a new story about AI “agents,” loosely defined as autonomous AI systems designed to navigate open-ended settings. 

First, Anthropic's Claude Code took coding by storm. Then, Moltbook, the social network where AI agents even tried to create religions (albeit with dubious veracity), gained notoriety.

Now, agents are proliferating out from software engineering across a wide variety of industries. 

The Arrival of Agents and the Collapse of Old Assumptions

Public discourse about AI has often incorrectly blurred the line between AI's capabilities—its ability to perform various tasks—and its autonomy.

Public discourse about AI has often incorrectly blurred the line between AI's capabilities—its ability to perform various tasks—and its autonomy.

The result meant that the rise of AI agents had often been overhyped as model capabilities grew. But now, the agents have firmly arrived.

These agents might have immense benefits. They could act as co-scientists to support scientific research, personalize education services, and even democratize software development.

AI agents also raise significant anxieties. Unexpected interactions among multiple agents can cause issues such as flash crashes or other unwanted behaviors. Agents themselves might be attacked by other untrustworthy entities. 

Public fears grew that Moltbook's agents might trigger an uprising, even if they were not well-founded. Yet despite these concerns, many governments and companies—many of whom already struggle to keep pace with AI—remain unprepared for agents.

The result risks the worst of both worlds: failing to mitigate harms while capturing no benefits.

We believe there are three key steps worth exploring, with varying roles for governments and companies across each step. 

So what should be done? No one has all the answers, but we believe there are three key steps worth exploring, with varying roles for governments and companies across each step.

1. Building the Missing Layer: Trust Infrastructure

Any ecosystem of AI agents may require a trust infrastructure to ensure its security. Building that ecosystem requires technical infrastructure for AI agents, ranging from IDs modeled after digital certificates to verify agents to mechanisms to correct harmful agent behavior. 

The infrastructure needed may vary depending on the agent's purpose, the risk of adversarial attacks, and other factors. Spreading such infrastructure, however, would not be easy, as it requires different companies to build and use agents to buy in. Therefore, we believe firms will lead technical infrastructure development, playing a central role in tasks like assigning IDs.

Given the challenges of coordinating such entities, we also believe public standards-setting organizations like the National Institute of Standards and Technology (NIST) should step up to coordinate public-private partnerships to institutionalize trustworthy infrastructure. 

Public procurement requirements or other mechanisms could be effective ways to incentivize the adoption of infrastructure standards. 

2. Governing Multi-Agent Systems

It's worth preparing for multi-agent interactions. On their own, AI agents can be quite impactful, but some of their most important interactions might involve working with multiple AI agents, including preparing for potentially unintended consequences. 

For example, the literature has highlighted that AI agents intended to protect car passengers may struggle to coordinate their strategies depending on the data on which they are trained. This is an area where both governments and companies could step up together.

These entities may want to support greater research on multi-agent systems and to promote international engagement to develop shared protocols for responding to unintended multi-agent behaviors.

3. Building Capacity and Resilience for an Agentic World

Beyond technical infrastructure, building broader societal capacity to both beneficially use and build resilience against AI agents would be valuable.

On the capacity side, this might involve helping individuals adapt to an agentic world. For example, students should receive better education on how to harness agentic coding tools to enhance their productivity. 

This may also involve social and legal infrastructure for agents, such as clear liability regimes, so individuals and firms can use and adopt them responsibly. Governments, in turn, could play a critical role here, as many of these reforms, such as to education or liability, require public action. On the resilience side, building clear muscle for responding to harmful agentic actions may be valuable as well. 

For example, if agents spread harmful deepfakes, we need to invest in public awareness campaigns or content provenance measures to mitigate them. Economists may want to better analyze concerns around agents and job displacement. 

Experimentation as Governance: Why Governments Must Act Now

While these three steps may require companies and governments to work together, government participation in any form within the agent ecosystem would be a powerful way to both build public-sector capacity and exert influence within it. 

Government participation could include experimenting with agents across several use cases, from automating administrative burdens that limit enrollment in public benefits to enhancing cyber defenses for digital infrastructure. 

Ideas akin to Tiago Peixoto's proposal for a public option for agents could be instrumental to ensure that all citizens have access to valuable agents.

More broadly, experimentation, which does not mean governments need to uncritically adopt agents for every task, can offer a compelling way for the state to identify what agentic use cases truly serve their citizens.

Experimentation can offer a compelling way for the state to identify what agentic use cases truly serve their citizens.

Government experimentation should include going beyond simply providing AI agents to public officials for well-known, specified use cases, such as data analysis. 

Rather, governments’ use of agents should focus on institutional experimentation, identifying unique public-sector needs where private-sector agentic practices may not yet exist. For example, developing security standards for code written by agents and used in sensitive domains, such as national security. 

This institutional experimentation should, in turn, be housed in specialized agencies that can look across government and identify unique opportunities, as exemplified in agencies like the UK Government's Incubator for AI. 

 Above all, however, we believe governments need to begin such experimentation now.

Above all, however, we believe governments need to begin such experimentation now. Agents are already here and will likely improve, making new best practices for their use in these domains vital. 

Together, this kind of experimentation might help shape the broader agentic ecosystem for the better.