Brookings: AI Agents Demand Governance That Acts Continuously and at Speed

Fast gains in capability and market size

The article, written by Landry Signé of Brookings' Center for Technology Innovation, points to a sharp acceleration in agents over the past two years. In 2025, agents improved their accuracy in computer tasks from 12% to 66.3%, which leaves them behind the average human baseline by just 6 points.

On market size, it cites Precedence Research, which estimates growth from $8 billion in 2025 to $199 billion by 2034. According to the author, this pace implies that existing governance models for AI more generally might not be sufficient.

The piece draws on a May 2026 World Economic Forum panel at the AI+Expo in Washington, D.C. A key theme there was that risk-based approaches, life-cycle governance, accountability, safety and human oversight remain relevant but must become far more operational as systems act faster and with less direct human intervention.

Where agents are already being used

Brookings lists uses across productivity, public services, education, health care, translation, accessibility, growth of small- and medium-sized enterprises and development goals in both the Global North and the Global South. Agents are already compressing research timelines for pharmaceutical companies, optimizing industrial goods supply chains and reducing customer service costs.

In manufacturing, agents have moved from flagging a defect to analyzing the systems, finding the probable cause and removing the defective unit. The article says these agents do not require large models or the newest chips, which makes them more accessible to lower-income countries than full digital transformations.

A World Economic Forum study of 70 core government functions found public services among the highest-readiness targets for agents. It cautioned that information-driven tasks should come before judgment-based ones such as eligibility assessment, where agents would affect sensitive decisions about individuals. The U.S. Centers for Disease Control and Prevention estimates a 527% return on its staffwide AI chatbot.

The author adds a caveat: documented evidence of deployed value is largely pre-agentic, so the governance question is arriving ahead of the deployments that will test it. Adoption is dominated by knowledge workers in higher-GDP countries, and enterprise uptake trails consumer use.

Why autonomy changes the risk picture

Agents combine autonomy with authority, permissions and access rights that let them act within defined boundaries. Existing governance is built on discrete stages, from model development to procurement, deployment and output review. Agents blur those boundaries because they pair models with memory, orchestration, tools and access to external systems.

The article contrasts them with traditional chatbots, which can be reviewed by humans before action is taken. Agents, by design, can invoke another system or action before a human intervenes. The risks it names include goal misalignment, behavioral drift, tool misuse and coordination failure. In this sense, agents start to resemble delegated decisionmakers and need much stronger oversight.

Even a properly authorized agent can fail, for example by misunderstanding a task or failing to recover after an unexpected event. A model may be low risk when retrieving information but higher risk when modifying records, communicating or initiating transactions.

Proposed tools and open problems

Brookings suggests assessing at least five interacting factors: capability, authority, architecture, context and consequence. It highlights the Agent Capability and Authorization Profile developed by the World Economic Forum. Agent cards document what an agent can do, while the profile records what it has been authorized to do, making scope, permissions, controls, evaluation evidence, monitoring and ownership explicit for a specific deployment. The article says these records should complement, not substitute, technical evaluation of the whole system.

Security is a major concern. Prompt injection is described as the most common attack technique, and because agent architectures treat all inputs as a unified token sequence, it is hard to identify which part was interfered with. That limits traceability and verification. Data used to power agents could also leak.

New protocols such as the model context protocol and agent-to-agent protocol let agents interact across platforms without needing to integrate them. This promises productivity gains but, the author says, requires more intentional governance because attacks could do more damage with greater access to connected systems. The article also flags accountability gaps between developers, deployers, users and affected communities.

What it means

The takeaway for organizations and policymakers is that stage-based reviews tied to design, procurement or deployment will not be sufficient as deployment decentralizes and autonomy rises. Because every step in an agent's chain of action accumulates new risk, Brookings argues that governance must act continuously and at speed. The source text available here ends partway through its discussion of the opportunity for smaller, multilingual, open or reusable models, which it says can support translation and accessibility with substantially less compute.