The agent economy needs more than agents

AI agents are usually described by what they can do.

They can research a market, write software, contact a customer, reconcile an account, arrange travel, negotiate a purchase or coordinate other agents.

That is the visible part of the market. It may not be where the largest or most durable value is created.

The moment an agent moves from producing information to taking consequential action, it encounters the same questions that surround every participant in an economy.

Who is it? Who does it represent? What is it allowed to do? Which systems can it access? Can its claims be trusted? Can it spend money? Can it make a commitment? Who is accountable if it is wrong? What evidence exists when two parties disagree?

The developing agent economy therefore needs more than intelligence. It needs rails, controls and institutions.

From software user to economic actor

Consider a company instructing an internal procurement agent:

Find a suitable security-testing provider for our new product. Spend no more than £15,000. Use an approved supplier where possible. Do not agree to a contract longer than 30 days. Escalate anything involving customer data.

The instruction is easy to express. Executing it safely is not.

The agent must establish which company it represents and demonstrate that authority to external systems. It needs access to an approved-vendor list, security policies and budget information. It must discover credible providers, compare capabilities and prices, communicate requirements, interpret terms, recognise when approval is required, record the basis of its decision and make or initiate a payment.

The company then needs to know what happened. Which options were considered? Which evidence was used? Who approved the purchase? Was the service delivered? Should the provider be paid? What happens if the outcome is disputed?

This is not a single agent task. It is an economic workflow.

Twelve layers around the agent

The infrastructure can be understood through twelve functional layers.

1. Intelligence and compute

Agents require models and inference capacity. They may need to select between providers according to quality, latency, cost, privacy and availability.

The model is important, but it is increasingly one component within a wider system.

2. Runtimes and orchestration

Agents need somewhere to plan, execute, retain state and coordinate. Companies using many agents need to know which agent owns which task, what happens when one fails and how work moves between them.

3. Tools and data access

Useful agents need access to business systems. That access must be authenticated, limited and observable. Connecting an agent to a CRM, bank account or customer database is fundamentally different from allowing it to search public information.

4. Identity and credentials

The receiving party must know which agent is making a request, which person or organisation it represents and whether those assertions can be trusted.

An identifier alone is insufficient. The economically useful question is the relationship between identity, representation and authority.

5. Authority and permissions

Agents should not receive unlimited permission merely because they act on behalf of an authorised company.

Authority must be bounded: this agent may issue a refund up to £200, contact these customer segments, use these data sources or purchase from these suppliers. Anything outside the mandate requires escalation.

6. Security and resilience

Agents consume untrusted information and call external tools. That creates opportunities for malicious instructions, compromised integrations, data leakage and unexpected behaviour.

Security must cover the agent's entire action path, not only the model endpoint.

7. Evaluation and observability

Companies must be able to test agents before deployment and observe them afterwards. Static benchmarks cannot answer every question about a live workflow.

Evaluation must include task completion, trajectory, tool use, policy compliance, cost, exceptions and changes in behaviour over time.

8. Discovery and procurement

Agents and companies need ways to find external capabilities. A useful registry must do more than list products. It must describe capabilities, evidence, commercial terms, security requirements and suitability for a defined task.

9. Contracts and outcome verification

Before one agent purchases work from another, the parties need a shared understanding of scope, price, authority, service level and acceptance.

Payment by outcome also requires a credible method of determining whether the outcome occurred.

10. Payments and financial rails

Agents need bounded ways to initiate and settle transactions. Companies need spending limits, approved counterparties, reconciliation, tax treatment and a clear record of which principal funded which action.

11. Governance and accountability

Management must know which agents exist, who owns them, what risks they create and whether their behaviour remains within policy.

Boards, auditors and regulators will require evidence that consequential actions can be explained and investigated.

12. Human supervision and workforce

Autonomy does not eliminate people. It changes where they enter the process.

Organisations need approval interfaces, exception queues, escalation routes and clear roles for the people supervising agent work. They also need to understand the total human labour hidden behind apparently autonomous systems.

Standards will matter. Standards will not be the whole market.

Protocols such as MCP and A2A are helping systems connect and agents coordinate. Payment and commerce initiatives are beginning to address mandates and settlement.

These foundations are important, but protocols often make interaction possible without resolving whether the interaction should be trusted.

The commercial value may accumulate in the control, evidence and workflow layers built above and around open standards: the systems that help a company decide which agent to use, what authority to provide, how performance will be measured and what happens when the work fails.

The second-order market

The first-order effect of better agents is that more tasks can be automated.

The second-order effect is that companies begin using many specialised agents across important workflows.

The third-order effect is that those agents require their own economic infrastructure: identity, permissions, procurement, reputation, payment, accounting, insurance, dispute resolution and governance.

That third-order market is still early. Many categories will be absorbed by existing enterprise platforms, cloud providers, payment companies and security vendors. Some apparently important layers will become features or free standards.

But some will become essential systems of control or networks of trust.

The central strategic question is not simply, “Which agents will win?”

It is:

Which institutions become necessary when agents are permitted to act?

That is the economy The Agentic Observer has been created to follow.