Issue Zero: The economy forming around the agents

Welcome to Issue Zero of The Agentic Observer.

This publication begins with a simple belief: the agent economy will be much larger than the market for agents.

The applications are the visible layer. Underneath them, an economic and institutional system is beginning to form—connections, identity, authority, security, evaluation, discovery, contracts, payments, governance and human supervision.

Those layers matter because the defining characteristic of an agent is not that it can generate fluent text. It is that software is being given some degree of discretion to pursue a goal and change the state of a system around it.

That raises a different class of question.

Not only: Can the agent do the work?

But also:

• Who authorised it?
• What is it allowed to access?
• Which actions can it complete without approval?
• How is its behaviour observed?
• What evidence exists after the action?
• What happens if it is wrong?

The Agentic Observer has been created to follow those questions—and the companies being built to answer them.

THE IMPORTANT DEVELOPMENT

Interoperability is becoming real infrastructure.

For several years, “agents working together” was largely a product demonstration or a framework-specific promise. The emerging protocol layer is making the idea more concrete.

The Model Context Protocol standardises how AI applications connect to tools and data. Its July 2026 specification introduced a stateless protocol core, a formal extensions framework, tasks and further authorisation hardening. The Agent2Agent Protocol provides a common interaction model through which independent agents can discover capabilities, communicate and delegate work. Google’s Universal Commerce Protocol addresses parts of the shopping lifecycle, while AP2 and x402 address different dimensions of agent-enabled payment. MCP specification, A2A specification, UCP overview, AP2 announcement, x402 documentation

This does not mean interoperability is solved.

It means the market can begin moving from bespoke connections towards shared ways of describing tools, capabilities, tasks and transactions.

That matters because a large ecosystem cannot form if every agent, tool, merchant and payment provider needs a unique integration with every other participant.

THE COMMERCIAL INTERPRETATION

Open protocols make the control layer more valuable—not less.

When connectivity becomes easier, more agents can reach more tools and counterparties. That increases utility. It also increases the number of actions that require policy, evidence and supervision.

A protocol may let an agent discover a tool. It does not decide whether this agent should be allowed to use that tool with this customer’s data.

A protocol may let two agents exchange tasks. It does not determine whether one party’s capability claim is credible, whether the price is fair or whether the result satisfies the buyer’s acceptance criteria.

A payment protocol may let an agent settle a request. It does not, by itself, answer whether the principal intended the purchase, whether the counterparty was approved or how the action should be reconciled in the company’s accounts.

Open standards therefore reduce one kind of friction while exposing another.

The valuable enterprise questions move upwards:

1. 1.Which agents exist?
2. 2.Who owns them?
3. 3.What identities and credentials do they use?
4. 4.Which tools and data can they access?
5. 5.What authority has been delegated?
6. 6.When must a person approve?
7. 7.How is performance tested and monitored?
8. 8.Which records prove what happened?
9. 9.How can authority be changed or revoked?
10. 10.Who handles failure, dispute or recovery?

The companies that answer those questions may become more consequential than many of the agent applications connected through the protocols.

COMPANY CATEGORY UNDER OBSERVATION

Agent evaluation and observability

The first generation of AI observability focused heavily on prompts, responses, latency, token consumption and trace inspection.

Agents expand the problem.

A useful trace may now include planning, retrieval, multiple tool calls, state changes, retries, hand-offs, policy decisions and human interventions. The unit of evaluation is no longer a single response. It is an entire trajectory through a business process.

The category includes platforms such as LangSmith, Braintrust, Arize, Langfuse, Weights & Biases Weave, AgentOps, Patronus AI, Giskard and Confident AI. Inclusion here is classification, not endorsement. Each company describes a different combination of tracing, testing, datasets, evaluation, monitoring and improvement. LangSmith, Braintrust, Arize, Langfuse, Weights & Biases Weave, AgentOps, Patronus AI, Giskard, Confident AI

The important buying question is not “Which platform has the best dashboard?”

It is:

Can this system help us decide whether the agent completed the intended business task, stayed within policy and remains safe to operate as its models, tools and environment change?

The strongest evaluation systems will connect technical traces to business outcomes and control obligations.

DEPLOYMENT FILE

The minimum safe path from demonstration to production

Suppose a company wants an agent to handle inbound sales enquiries.

The tempting route is to connect it to the website, CRM, calendar and knowledge base, then judge success by meetings booked.

The safer and more informative route is staged.

Stage one: replay. Give the agent historical enquiries and compare its proposed actions with what skilled people did. Do not permit external action.

Stage two: shadow. Let the agent observe live enquiries and recommend a response, qualification status and next step. A person still acts.

Stage three: constrained action. Permit the agent to answer only defined low-risk enquiry types, use approved sources and book meetings within bounded rules.

Stage four: exception-led operation. Expand authority only where evidence supports it. Route uncertainty, policy exceptions, high-value accounts and unusual requests to people.

Stage five: monitored production. Continue sampling, incident review, performance comparison and authority review. Production is not the end of evaluation.

The decision at each stage should be evidence-based: deploy, remediate or reject.

Our complete controlled-pilot playbook is now available in Deployment Files.

INFRASTRUCTURE WATCH

Payments are becoming an identity-and-authority problem.

x402 allows an HTTP resource to quote a price and receive payment inside the same request flow. Stripe’s Agentic Commerce Suite is designed to help merchants make products discoverable and accept purchases initiated through agents. Google’s AP2 introduces mandates intended to provide evidence of user intent and authorisation. Mastercard’s Agent Pay for Machines describes credentialing, permissioning and settlement for high-frequency machine transactions. x402, Stripe Agentic Commerce Suite, Google AP2, Mastercard Agent Pay for Machines

The common thread is important.

Agent payments are not merely faster checkouts. They require a reliable connection between the principal, the agent, the mandate, the merchant and the resulting financial record.

The winning infrastructure will not only move money. It will make delegated spending understandable and controllable.

FIVE SIGNALS

1. NIST has made agent standards an explicit area of work.

The US National Institute of Standards and Technology launched an AI Agent Standards Initiative in February 2026 focused on confidence, security and interoperability. Standards activity does not prove near-term enterprise adoption, but it is evidence that agent identity, secure action and interoperability are becoming institutional questions rather than isolated product features. NIST AI Agent Standards Initiative

2. MCP is moving towards enterprise readiness.

The 2026 MCP roadmap highlighted transport scalability, agent communication, governance maturation and enterprise readiness. Watch the authorisation model, registry/discovery mechanisms and the boundary between open protocol and commercial gateway. MCP 2026 roadmap

3. A2A 1.0 formalises a cross-vendor interaction layer.

A2A’s move to a 1.0 specification gives builders a more stable basis for cross-agent discovery and delegation. The next question is how capability claims, identity and trust will operate across organisational boundaries. A2A 1.0

4. Commerce protocols are converging on existing infrastructure.

UCP is designed to work with existing retail systems and multiple transports. Stripe positions its suite as protocol-agnostic. This suggests merchants will prefer abstraction over betting their commerce roadmap on one standard. UCP, Stripe technical field guide

5. Discovery is becoming a commercial layer.

A2A uses Agent Cards to describe capabilities; Coinbase’s x402 Bazaar makes paid services searchable; cloud marketplaces are adding agents. Discovery will become more useful when capability descriptions are connected to credible performance, security and commercial evidence. A2A agent discovery, x402 Bazaar

THE VERDICT

Buy: Evaluation and control around a clearly defined, high-value workflow.

Test: Protocol-based connections where they reduce integration work without expanding permissions prematurely.

Monitor: Agent discovery, payment mandates and service procurement across organisational boundaries.

Avoid: Treating autonomy as a product feature that can be switched on before authority, evidence and recourse have been designed.

The agent economy will not arrive in one launch.

It will form each time an organisation gives software another permission, another tool, another budget or another decision.

Our job is to observe what becomes necessary as that happens.