By Ramesh Raskar, Navin Chaddha, Vijay Reddy, Rahul Todkar

AI developers, companies and users face a basic but unresolved question: How much autonomy should an AI agent receive? Today, systems that require approval for every action and systems that can plan, adapt and execute for hours are all described simply as “agents,” leaving us without a meaningful way to specify human supervision or accountability.

The autonomous-vehicle industry faced a similar problem. It responded with the L1–L5 framework, which describes how responsibility shifts from the human driver to the automated system. AI needs an equivalent A1–A5 framework—one that connects increasing autonomy with decreasing real-time human supervision and stronger forms of governance.

At the lower levels, safety comes primarily from keeping a human in the loop. As autonomy increases, safety must move into the agent’s policies, operating environment and governing institutions.

From Driving Intelligence to Agent Intelligence

The analogy between autonomous vehicles and autonomous agents is not exact, but it is useful.

The AI model provides both the engine and the driving intelligence. It supplies the reasoning, perception, planning and decision-making capabilities. The agent is the complete vehicle: the model combined with memory, goals, credentials, tools and the ability to act.

The internet provides the existing roads. But those roads were designed primarily for humans, websites, applications and APIs—not for autonomous agents that discover services, negotiate with counterparties, spend money and delegate work to other agents.

The emerging Agentic Web must therefore provide the equivalent of lanes, maps, traffic rules, licenses, registration, insurance and accident investigation. It must determine which agents may enter the road, whom they represent, where they may operate and who is responsible when something goes wrong.

The automotive analogy also helps resolve a common confusion. A vehicle can have some automation without being fully autonomous. Similarly, an AI system can be an agent without being an independent agent.

All A1–A5 systems are agents. What changes is how much real-time human supervision they require.

A model or assistant that only produces information, without authority to take action, can be treated as A0. The A1–A5 ladder begins when the system can act.

A1: Directed

At L1, a vehicle provides limited assistance. It may help control either steering or speed, but the human remains fully engaged. Hands, eyes and mind remain on the driving task.

An A1 Directed agent performs individual actions under immediate human direction. The person defines the task, selects the relevant tool and approves consequential actions.

An A1 travel agent may search for flights but require the user to choose the itinerary and approve the purchase. An A1 financial agent may prepare a payment but be unable to submit it. An A1 coding agent may propose a change but require approval before modifying the repository.

The human is still driving. The agent is assisting.

Governance at A1 is relatively direct: restricted tools, visible actions and explicit approval. The central requirement is that the system not quietly convert assistance into execution.

A2: Delegated

At L2, the vehicle may control steering and speed together, but the driver must continue monitoring the road. Some systems may permit the driver to remove hands from the controls, but not eyes or attention from the driving task.

An A2 Delegated agent receives a bounded objective and carries out multiple steps without asking for approval at every stage. The human defines the operating envelope: the tools the agent may use, the data it may access, the amount it may spend and the duration of the delegation.

An A2 travel agent might reserve a flight within specified dates, airlines and a $1,000 limit. An A2 procurement agent might reorder approved supplies from an existing vendor within a predefined budget.

The person is no longer controlling every movement, but remains responsible for supervising the assignment.

A2 governance requires scoped credentials, approved tools, spending limits, destination restrictions and automatic expiration. Giving an agent a task should not mean giving it all of the user’s permissions. Delegation must be specific, bounded and reversible.

A3: Adaptive

L3 marks a fundamental transition. Under defined conditions, the automated system performs the complete driving task and monitors the environment. The human need not watch continuously but must remain available to take over when requested.

An A3 Adaptive agent can revise its plan as circumstances change. It can choose among tools, recover from ordinary failures and respond to routine exceptions. When it encounters a situation outside its operating envelope, however, it must stop or escalate to a human.

Suppose a travel agent discovers that the selected flight has been canceled. An A2 agent may simply report failure. An A3 agent can search for alternatives, compare trade-offs and rebook within the original constraints. But if every alternative requires changing the destination, exceeding the budget or sharing personal information with an unfamiliar service, the agent must request intervention.

A3 creates the agent equivalent of the autonomous-vehicle handoff problem: Can the system recognize that it has reached the limits of its competence or authority early enough for a human to intervene safely?

Governance at A3 requires continuous trajectory monitoring, reliable escalation, preserved provenance and safe stopping. The agent must know not only how to continue, but when it should not continue.

A4: Self-Governing

At L4, the automated-driving system performs the driving task and handles failures without relying on a human fallback—but only inside a defined operational domain. That domain might be a mapped city, a designated highway or a limited set of weather conditions.

An A4 Self-Governing agent operates without routine human supervision inside a similarly defined agent operating domain.

That domain might specify:

• The organizations and services the agent may contact
• The tools, models and data it may use
• Its financial and contractual limits
• The jurisdictions in which it may operate
• The policies it must enforce
• The conditions under which it must stop
• The records it must produce

An A4 procurement agent might negotiate with approved suppliers, place orders within budget, monitor delivery and resolve routine disputes without a person watching each transaction.

At A4, human oversight moves from supervising actions to governing the system. People establish policies, certify the agent for a particular domain, audit its behavior and investigate incidents.

This level requires machine-enforceable policy, persistent identity, complete action receipts, revocable credentials, independent monitoring and defined liability. The agent must be capable of reaching a safe state without assuming that a human will always be available to rescue it.

A5: Independent

At L5, a vehicle can perform the complete driving task across the range of roads and conditions that a competent human driver could handle. The occupants are passengers, not fallback drivers.

An A5 Independent agent can operate across organizations and changing environments. It can discover services, evaluate counterparties, negotiate terms, coordinate with other agents and complete extended tasks without operational human supervision.

An A5 personal agent might communicate with healthcare providers, banks, retailers and government agencies. An A5 business agent might identify suppliers, negotiate contracts, arrange financing and coordinate logistics across jurisdictions.

At this level, safety cannot depend on one user watching the agent or one company controlling the entire environment. Governance must move to the ecosystem.

A5 agents will need verifiable identities, portable reputations, standardized delegation credentials, contractual protocols, dispute-resolution mechanisms, auditability, revocation and clear legal responsibility. Their interactions must be governed by shared rules of the road rather than by the policies of a single platform.

But independent operation must not be confused with unlimited power.

A5 means independent execution, not independent authority.

A fully automated vehicle may drive without human assistance, but it still cannot enter a closed road, ignore traffic laws or cross a national border without authorization. Similarly, an A5 agent should remain bound by the authority of the person or organization it represents.

Autonomy describes how independently an agent can operate. Authority determines where, when, how and for whom it may act.

Human Supervision Declines; Governance Must Increase

The progression from A1 to A5 is sometimes described as removing the human from the loop. That is incomplete. The human role does not disappear. It moves to a different level.

Level

Vehicle supervision

Agent supervision

Primary form of governance

L1 / A1

Hands, eyes and mind on

Human approves individual actions

Direct control

L2 / A2

Hands off; eyes and mind on

Human delegates but continuously supervises

Bounded delegation

L3 / A3

Eyes off; human must remain available

Agent operates independently but escalates exceptions

Conditional oversight

L4 / A4

Mind off; Human attention not required within the operating domain

Agent handles execution and failures within a defined domain

Embedded policy and certification

L5 / A5

No operational human fallback across domains

Agent operates across organizations and environments

Ecosystem governance

At A1, the human governs individual actions. At A2, the human defines and monitors the task. At A3, the human handles exceptions. At A4, the human establishes policy and certifies the operating domain. At A5, people create the institutions, laws and accountability mechanisms within which independent agents operate.

As real-time human supervision declines, governance cannot disappear. It must migrate from the individual user into the agent, its operating environment and the institutions around it.

Intelligence Enables Autonomy—and Autonomy Compounds Intelligence

More capable models make greater autonomy possible. An agent that can understand context, construct plans, recover from failures and recognize exceptions can operate with less immediate human supervision.

But the relationship also runs in the other direction.

Once an agent is allowed to operate, it gains access to tools, observations, feedback, memory and other agents. Its experience in the world can make it more effective. It can learn which plans succeed, which counterparties are reliable and which strategies work under different conditions.

Intelligence enables autonomy, and autonomy can compound intelligence.

This creates a powerful feedback loop. As agents become more capable, organizations will grant them more independence. As agents operate more independently, they will accumulate experience and become more capable.

Governance must therefore advance alongside both intelligence and autonomy.

Greater capability should not automatically produce greater authority. An agent should move from A1 to A2—or from A3 to A4—only when the surrounding controls, monitoring and accountability mechanisms are ready to support that transition.

A highly intelligent model may deliberately remain at A1 because the consequences of an error are severe. A less sophisticated agent may operate safely at A4 inside a narrow and carefully controlled environment.

The appropriate autonomy level depends not only on intelligence, but also on risk, reversibility, operating domain and the maturity of the surrounding governance.

Governance Makes Autonomy Possible

The purpose of A1–A5 is not to declare A5 the goal for every agent. Some systems should remain at A1 permanently. Others may be safe at A4 within a narrow operating domain but unsafe at A5.

The goal is appropriate autonomy.

Before an agent advances from one level to the next, it should demonstrate not only improved task performance but also reliable operation under the governance conditions required at that level.

Every deployed agent should make several facts visible:

• Its autonomy level
• The principal it represents
• Its operating domain
• The authority it has received
• The tools and data it can access
• Its financial and temporal limits
• Its escalation and safe-stop behavior
• The party responsible for its actions

Higher-level agents will also require conformance testing, adversarial evaluation, incident reporting, action receipts, revocation and clear liability.

The governing principle should be:

No agent should operate at an autonomy level higher than the governance environment around it can safely support.

Transportation did not become scalable because every vehicle became perfectly safe. It became scalable because vehicles evolved together with roads, standards, licensing, insurance and law.

AI agents will follow the same pattern. A1 and A2 agents can rely heavily on human supervision. A3 agents require reliable handoffs. A4 agents require enforceable operating domains. A5 agents require a functioning Agentic Web.

Governance is not what prevents autonomy. Governance is what makes greater autonomy possible.