The agentic economy market map

On this page

Chapter 3 of The Hitchhiker's Guide to the Agentic Economy

Report hub · Previous: AI agent adoption and measurement · In this chapter: agent systems · commerce · actors and loss · service snapshot · enterprise boundary

Our view

The agentic economy is a chain of interdependent markets organised around delegated work. Models supply reasoning capability, but useful action also depends on data, runtimes, tools, identity, policy, evaluation, security, discovery, commercial state, payments, enterprise systems and accountable organisations. For a buyer, the map answers four questions: what job is moving, what can the software touch, how does money move, and who fixes the result?

An agentic economy market map is a functional view of the suppliers, systems and institutions needed to move one delegated job from objective to accepted result.

Agentic Economy uses four views to organise this early market: work and demand, agent systems, exchange and commerce, and institutions and consequences. This is a working synthesis, not an official industry taxonomy. Five flows join the views: work, data, authority, money and accountability. The gaps between layers are where many products are forming and where most production failures appear.

Key Takeaways

• The market has four connected systems and more than a dozen distinct infrastructure layers.
• APIs, MCP, A2A and commerce protocols solve adjacent problems; they are not direct substitutes for every use case.
• Commercial completion requires discovery, authority, transaction, delivery, acceptance and records.
• Existing enterprise and public infrastructure remains part of the stack.
• Market value is likely to collect around scarce capability, trusted context, distribution, authority, systems of record and accepted outcomes.

Sources and scope

Agentic Economy built this functional map from primary protocol specifications, provider documentation, official Australian sources and the dated service checks cited below. The taxonomy classifies a participant by the job it supplies, not its marketing label. The map and five-flow model are Agentic Economy analysis dated 13 September 2026. About Agentic Economy explains the publication. Readers can submit corrections through the contact page.

The market has four connected systems

We organise the agentic economy into four connected systems: work and demand, agent systems, exchange and commerce, and institutions and consequences. This is not a vendor stack. It is a way to place companies according to the role they perform and the boundary they control.

The same company can occupy several layers. Model providers are moving into tools and agents. Cloud platforms are combining inference, identity, runtime and observability. Payment companies are extending credentials and checkout systems to software buyers. Enterprise incumbents are adapting their systems of record rather than yielding them to a new agent-only stack.

The market in one diagram

                          INSTITUTIONS AND CONSEQUENCES
          law | standards | public infrastructure | labour | capital
                                      |
                                      v
WORK AND DEMAND                AGENT SYSTEMS                 EXCHANGE AND COMMERCE
objective                      models + compute              discovery + offers
workflow        ----------->   context + memory   -------->  checkout + orders
buyer and user                 runtime + tools               credentials + payments
outcome                        identity + policy             delivery + acceptance
repeat use      <-----------   evaluation + security <----  billing + recourse
                                      |
                                      v
                         ENTERPRISE SYSTEMS OF RECORD
                ERP | CRM | finance | procurement | identity | audit

Across every connection: work | data | authority | money | accountability

The diagram is deliberately circular. Demand gives the system a job. Agent infrastructure interprets and performs it. Exchange connects buyers and suppliers. Institutions establish the rules and consequences. Enterprise systems retain the durable records. Evidence from the completed job returns to the operator, supplier and market. The web edition presents this as an accessible SVG; the text and tables remain the complete alternative description.

View one: work and demand

The market starts with a job that somebody values. Without that job, agent capability is supply looking for a problem.

Layer	Function	What to observe
Objectives and tasks	Convert a human or organisational need into bounded work	Clear deliverable, stopping condition and consequence
Workflows and sectors	Place the task inside an operating process and industry context	Prior process, hand-offs, systems and constraints
Buyers and users	Identify who pays, who operates the system and who receives the benefit	Budget owner, user, beneficiary and accountable principal
Outcomes and repeat use	Establish whether the work was useful and durable	Acceptance, time, cost, quality, intervention, retention

Current demand is visible in software development, customer support, research, marketing, sales operations, finance, procurement and administrative work. The evidence varies sharply. A provider can report usage of a coding agent or support feature. An enterprise can report a production deployment. A buyer can describe time saved. Those records become economically comparable only when the task and outcome are defined.

This view also prevents a common category error: high generative-AI use does not automatically mean high agentic adoption. The relevant event is delegated action inside a workflow. Chapter 2's job-level measurement record provides the minimum evidence for that claim.

View two: agent systems

Agent systems turn an objective into controlled action. The market includes the intelligence layer and the software required to operate it over time.

Layer	Job supplied	Representative participants or approaches	Market question
Models and inference	Reasoning, language, vision, planning and tool selection	OpenAI, Anthropic, Google, Meta, Mistral, Cohere; open model ecosystems	Which capability, latency, cost and deployment boundary fit the task?
Chips, networks, data centres and energy	Physical capacity for training, inference and connectivity	NVIDIA, AMD, cloud providers, data-centre and energy operators	Which capacity, geography and energy constraints shape cost and availability?
Compute and hosting	Deployment, inference, scaling and managed infrastructure	AWS, Microsoft Azure, Google Cloud, specialist inference providers	Where does compute scarcity or price constrain the service?
Context, memory, data and provenance	Retrieval, state, licensed inputs, source history and task-specific information	Databases, vector stores, search and data providers, knowledge platforms, data licensors	Is the information current, permitted and attributable?
Runtimes and orchestration	Agent loops, durable execution, hand-offs, scheduling and recovery	OpenAI Agents SDK, Google ADK, LangGraph, Microsoft Agent Framework, AWS Bedrock AgentCore	What state survives a long-running or failed job?
Tools, browsers and connectors	Access to software, websites and external services	APIs, browser-use providers, SaaS connectors, MCP servers, code execution	Can the agent complete the action reliably and safely?
Agent coordination	Discovery, delegation, communication and task state between agents	A2A and application-specific coordination systems	How do independent systems exchange tasks and evidence?
Identity, authority and policy	Principal identity, permission, mandate, budget, approval and revocation	Enterprise identity, secrets and policy systems; AP2 and payment-network controls	Who may do what, for whom, under which constraints?
Evaluation and observability	Testing, traces, quality measurement and incident investigation	Model evaluation, agent testing, tracing and monitoring providers	Can the organisation detect and explain failure?
Security and recovery	Containment, data protection, abuse defence and restoration	Sandboxes, security platforms, policy controls, human escalation	How is the blast radius constrained and the job recovered?
Implementation and human operations	Integration, process redesign, review, exception handling and managed delivery	Consultancies, systems integrators, BPO providers, domain specialists and internal operations teams	Who performs the work that remains outside the software boundary?

This is a functional map, not a final category assignment. Large vendors span several rows. Open-source projects can provide components used inside commercial products. An orchestration framework may include tracing; a cloud runtime may supply identity; a model platform may host tools. The report will classify each product by the job it performs and record overlaps explicitly.

The interfaces between agent layers

The acronyms describe different jobs. One connects a model to a tool; another passes a task between agents; another structures an order or payment. Supporting several is normal because none is the whole stack.

Several widely discussed interfaces sit at different boundaries:

Interface	Primary problem	Connects	Does not by itself establish
API	Expose a defined software operation	Application to application	Agent discovery, delegated authority or commercial terms
MCP	Give an AI application a standard way to connect to tools and context	Host, client and MCP server	Supplier trust, payment, delivery acceptance or enterprise procurement
A2A	Discover agent capabilities and exchange tasks, messages and artefacts	Agent client and remote agent	Shared internal state, tool internals, commercial authority or settlement
ACP or UCP	Express merchant capabilities, checkout and order state	Agent platform, buyer and merchant	Universal identity, every payment rail or post-sale operating process
AP2	Express mandate and payment-authority semantics	Principal, agent, merchant and payment participants	Fulfilment quality, accounting or legal agency
x402 or MPP	Initiate or settle machine payments	Software buyer, supplier and payment or wallet infrastructure	Delivery, acceptance, tax, refunds or durable supplier terms
Cards, wallets and network tokens	Supply payment credentials, authorisation and network processing	Payer, merchant, issuer, acquirer and network	Permission to perform the underlying business job

MCP and A2A are often described together because both support interoperability. Their roles are complementary. MCP commonly exposes a tool or context source to an AI host. A2A describes collaboration with another agent that can manage its own task. Google's protocol guide draws this distinction directly (Google Developers, "Developer's Guide to AI Agent Protocols"); MCP's own documentation defines its client-server architecture and capabilities (Model Context Protocol documentation).

Chapter 12 adds version or commit, owner, roles, transport, authentication model, state machine, retry and idempotency behaviour, implementation evidence, checked date and stopping boundary. This table remains a functional comparison rather than a conformance finding.

View three: exchange and commerce

Exchange turns capability into a buyer-supplier relationship. The market includes consumer product discovery, software service calls, machine payments and established enterprise commerce.

Layer	Function	Current approaches	Unresolved seam
Discovery and representation	Make a product, service or agent findable and comparable	Search, assistants, catalogues, agent cards, marketplaces, machine-readable product data	Who controls ranking, evidence and access to the customer?
Offer and quote	Define the job, inputs, output, price, timing and conditions	Product feeds, service schemas, capability negotiation, direct API terms	How are changing scope and professional judgement represented?
Checkout and order	Create and update the commercial transaction	ACP, UCP and provider-specific checkout interfaces	Which party owns order state, tax, cancellation and fulfilment?
Authority and credential	Prove the agent may transact under constraints	AP2 mandates, shared payment tokens, network agent tokens, wallet approvals	How does organisational delegation cross agents and providers?
Payment and settlement	Transfer value through an appropriate rail	Cards, bank payments, wallets, x402, MPP and stablecoins	Which rail fits value, frequency, geography, risk and recourse?
Delivery and acceptance	Return the product, service result or state change and validate it	Merchant fulfilment, API response, service workflow, human review	What proves that the customer's job succeeded?
Billing and finance	Meter, invoice, collect, reconcile and report	Subscriptions, usage billing, invoicing, ERP and accounting systems	How are agent events joined to commercial and tax records?
Support and recourse	Handle correction, refund, dispute and renewal	Provider support, network disputes, contractual remedies, human escalation	Who absorbs the cost of agent error across organisations?

Consumer checkout, machine payment and enterprise commerce

Three commercial patterns are forming at once.

Consumer agentic commerce lets an assistant discover or compare products and helps a person complete a purchase. Current work by OpenAI, Google, Shopify, Stripe, PayPal and card networks sits largely in this frame. The merchant commonly retains pricing, fulfilment and post-purchase responsibility while the assistant becomes a new discovery and interaction surface (Universal Commerce Protocol; Agentic Commerce Protocol).

Machine payments let software pay for a bounded digital resource or service, often at the point of use. x402, MPP and related products reduce the friction of accounts, API keys or pre-funded balances for some flows. Their public specifications and endpoints make this segment unusually inspectable. Payment activity still needs buyer, service, delivery and independence context.

Enterprise agentic commerce is agent-initiated work inside a durable buyer-supplier relationship. The transaction may sit under a supplier account, master agreement, purchase order, usage commitment, invoice and finance workflow. Existing billing, procurement, identity and accounting systems carry much of this state. Agentic interfaces can improve discovery, negotiation and execution while preserving the commercial control plane.

For example, an agent can buy one search result through a machine-payment endpoint without creating a durable supplier relationship. A recurring research service may instead need an account, authorised users, a usage commitment, monthly invoices, service levels, data terms and a correction process.

These patterns will overlap. A consumer wallet may issue a scoped credential to an agent. A business agent may use a card for a low-value purchase. A machine-payment protocol may settle a digital input inside an enterprise workflow. The correct architecture depends on the principal, frequency, value, supplier relationship, geography and consequence.

Follow the actors and the loss

The word agent can hide the people and institutions carrying the commercial risk. A complete transaction may involve this chain:

principal -> orchestrator -> agent -> tool or service -> merchant or provider
          -> payment participants -> delivery -> acceptance
          -> finance and reconciliation -> correction or dispute

The principal is the person or organisation on whose behalf the system acts. The user may request the work without being the payer or beneficiary. Payment participants can include a wallet or custodian, facilitator, acquirer, issuer, bank or stablecoin issuer. A reviewer may accept the result, while finance owns reconciliation and a supplier or insurer bears a later remedy.

At every hand-off, record five facts: who authorises the action, which system is the source of truth, who receives money, who bears the failure cost, and who can reverse or revoke the action. An economic principal is the person or organisation whose authority, funds and obligations sit behind the action. An agent is not the legal or economic principal merely because it acts on a principal's behalf.

View four: institutions and consequences

Agentic systems operate inside existing organisations and societies. The surrounding institutions determine which actions are legitimate, who remains responsible and how value is distributed.

Layer	Role in the market	Questions for the report
Enterprise systems	Retain customer, supplier, financial, procurement, identity and audit records	Which system is authoritative when an agent acts?
Firms and operating models	Allocate work, authority, management and liability	Which tasks move, which roles change and where does human judgement remain?
Labour and skills	Supply expertise, review, exception handling and redesigned work	Does the system substitute, complement or create tasks, and for whom?
Standards and governance	Define interoperability, risk practices and common controls	Which rules are specified, adopted, enforceable or still contested?
Law and regulation	Establish duties, rights, liability and sector obligations	Who is legally responsible for data, advice, transactions and harm?
Public infrastructure	Provide payment, identity, invoicing and other shared systems	How can agents use existing national infrastructure safely?
Capital and competition	Fund companies and influence market structure	Which layers concentrate, commoditise or retain pricing power?
Insurance and assurance	Price, transfer or certify defined risks	Which losses are covered, excluded, audited or left with the buyer?
Professional services and human operations	Integrate systems and supply judgement, review and exception handling	Is hidden labour part of the product cost and quality model?
Physical systems and robotics	Carry agent decisions into warehouses, vehicles, sites and devices	Which safety, latency, connectivity and maintenance boundaries apply?

Australia's institutional layer

Australian deployments use global technology while remaining subject to Australian obligations and infrastructure.

• Privacy: OAIC guidance recommends due diligence, transparency and human oversight when organisations use commercially available AI products (OAIC, privacy and commercially available AI products).
• Cyber security: ASD guidance focuses on the harness around the model, including tools, memory, permissions and exposure to untrusted content (ASD, agentic AI harnesses).
• Financial services: ASIC has highlighted governance gaps in AI use by financial-services licensees. Its REP 798 reviewed 23 licensees and 624 use cases as at December 2023, so the findings should not be generalised to every Australian business (ASIC, REP 798).
• Payments: PayTo can encode authorised terms for one-off, ad hoc and recurring account-to-account payments (Australian Payments Plus, PayTo).
• Business documents: Australia's Peppol eInvoicing framework exchanges structured invoice information between buyer and supplier accounting systems (Australian Taxation Office, eInvoicing).

These systems do not create agent authority as a single package. They supply parts of a production route. An enterprise implementation must join them to the organisation's own identity, delegation, contract, acceptance and audit controls.

This edition treats Australia as an operating lens, not a measured national market. A defensible Australian atlas still needs supplier and deployment records by sector, firm size and geography; public-sector and procurement coverage; and local evidence from small service businesses. Regulatory references are general operating context, not legal advice.

The five flows that join the market

Layer maps are useful until a real job crosses them. Then the interfaces matter more than the boxes.

Work

The work flow carries the objective, plan, task state, hand-offs and delivered result through one stable job identifier across agent and business systems.

Data

The data flow carries instructions, context, source material, tool responses and records. Each participant needs a clear right to access, retain and transform that information. Provenance and freshness determine whether the action is defensible.

Authority

The authority flow links the principal to the system and constrains action. It includes identity, roles, permissions, mandates, spending limits, approvals, expiry and revocation. Authority may be enforced in several layers; the constraints need to remain consistent.

Money

The money flow carries quote, credential, authorisation, payment, settlement, invoice, credit and refund. The rail follows the commercial relationship: a low-value digital call and a recurring enterprise service need different controls even when both are initiated by agents.

Accountability

The accountability flow connects logs, evidence, validation, acceptance, incidents, correction and recourse. It makes the job inspectable across organisational boundaries. This is the flow most likely to reveal that a technically successful call did not produce a commercially complete result.

Controls at the joins

Authentication, business authority, payment authorisation and legal responsibility are separate questions. The minimum control record makes that separation visible.

Failure mode	Control evidence to request
Untrusted content redirects the agent	Trusted-instruction boundary, content isolation, tool allow-list and review threshold
A credential is used outside its purpose	Principal identity, audience, scope, expiry, proof of possession and revocation path
The agent takes an excessive action	Tool, data, supplier and spend limits; approval thresholds; policy decision log
A request is replayed or duplicated	Stable job ID, idempotency key, replay protection and reconciliation rule
One agent impersonates another	Mutual authentication, signed task identity and capability verification
Data leaks across providers	Data-flow record, retention terms, regional processing boundary and deletion evidence
Delivery is wrong or incomplete	Acceptance test, human escalation, correction window, refund or contractual remedy
The job fails after payment	Joined order, payment, delivery and support records with one accountable owner

A control claim becomes assurance only when the evidence card is complete:

Field	Required record
Owner and implementation point	Named accountable role and the system where the control is enforced
Test and data	Reproducible method, environment, identities, inputs and negative cases
Expected result and threshold	Pass/fail condition, severe-failure rule and tolerated residual error
Cadence	Pre-release, change-triggered, periodic and incident-driven review
Evidence	Logs, decisions, receipts, screenshots or reports retained for a stated period
Exception and escalation	Who may accept an exception, for how long, and who contains a failure

The threshold cannot be universal. It follows the value, reversibility and consequence of the delegated action.

Where companies may build value

The following are Agentic Economy's current market hypotheses. The source set does not yet establish durable pricing power, retention or margin in these layers.

Source of value	Why it may remain scarce	Strategic test
Distinct capability	Some tasks require better models, proprietary methods or specialist execution	Does the product improve accepted outcomes, not only benchmark scores?
Trusted context	Current, permissioned and domain-specific data is difficult to assemble	Does the supplier control access, provenance or workflow-specific knowledge?
Distribution	Assistants and marketplaces can shape which suppliers are considered	Can the company reach buyers without depending on one gatekeeper?
Authority and trust	Cross-organisation delegation and risk controls are hard to standardise	Does the product reduce risk or merely add another credential?
System of record	Durable operational and financial state is costly to replace	Does the agent layer strengthen or bypass the authoritative record?
Accepted outcome	Delivery, quality control and recourse combine technology with operations	Can the supplier prove completion, margin and repeat demand?

Open interfaces may reduce integration cost while distribution remains concentrated. Model capability may commoditise while domain data becomes more valuable. Existing enterprise vendors may extend their systems of record, while new companies may win a narrow workflow by delivering a better accepted outcome. Testing those hypotheses requires observed pricing, switching, customer retention, support cost and accepted-job margin rather than provider documentation alone.

The business models are equally varied: per-seat access, usage charges, subscriptions, transaction fees, payment spread, managed services, outcome pricing, data licensing and marketplace commissions. The important dividing lines are who bears integration and support cost, what drives switching, how concentrated distribution is, and whether gross margin improves as accepted volume grows.

What participant claims establish

Every company or protocol record in the report answers the same questions:

1. Which layer and job does it supply?
2. Who is the buyer or user?
3. What product is actually available?
4. How does it charge or capture value?
5. Which interfaces and dependencies does it use?
6. What evidence shows use beyond documentation or announcement?
7. Is it available and relevant in Australia?
8. What part of the customer job remains outside its boundary?

A company does not enter a category because its marketing uses agentic. A protocol does not cover a business state because it can transport an adjacent message. The functional job, implementation and evidence decide the classification.

A dated service snapshot

These three records show the difference between an available component and a completed customer job. Agentic Economy checked the public documentation and pricing surfaces on 7 September 2026. The routes were publicly documented and reachable; they were not purchased, delivered or accepted.

Service route	Buyer can purchase	Checked price	Access and evidence	Boundary
Tavily direct x402 search	One advanced web search with optional page content	0.01 USDC per call on Base	Provider-operated POST /search; provider documentation and machine-readable pricing checked	Search results still need source review and integration into a useful job
Browserbase direct x402 session	Browser connection time	0.03 USDC for 15 minutes; 0.12 USDC per hour	Provider-operated session creation route and x402 documentation checked	The buyer's agent must still operate the browser and validate the result
Firecrawl through StableEnrich	Search or scrape access through a third-party gateway	US$0.0252 per search; US$0.0126 per scrape, paid in USDC	Gateway documentation and public service listing checked	Gateway terms and response surface differ from buying Firecrawl directly

The full checks, request shapes and business examples are documented in Agentic Economy's Tavily, Browserbase and Firecrawl field notes. The prices are dated offer prices, not current promises, observed average spend or proof of customer demand.

The enterprise adoption boundary

Enterprise adoption turns on ten questions that cut across every layer of the market:

1. What exact job is delegated, and who requests, pays for and accepts it?
2. Which choices, systems, data and spend are inside the agent's authority?
3. Where is data processed, how long is it retained and which system is authoritative?
4. Can the vendor provide a stable job ID, logs, receipts and evidence of delivery?
5. What counts as first-pass acceptance, rework and final acceptance?
6. Who corrects, refunds or escalates a failed result?
7. What are the usage cap, service level, support and termination terms?
8. Which actions remain human-reviewed, and why?
9. How are order, payment, invoice, tax and finance records reconciled?
10. Which result ends the pilot, and which incident stops it early?

Headline price becomes comparable only when the surrounding supplier terms are visible:

Term	Question to settle
Service boundary	Which job, action, evidence and support obligation is actually supplied?
Price basis	Which currency, tax, minimum, usage, overage and renewal rules apply?
Data and region	Where do requests, logs, backups and support access travel, and under whose terms?
Reliability	Which service level, quota, support response and incident process is committed?
Change	How much notice applies to a model, protocol, price or product retirement?
Audit and portability	Which logs, records, configuration and customer data can be inspected and exported?
Failure and remedy	Who owns correction, service credit, refund, dispute, indemnity and residual loss?
Exit and fallback	How does the buyer terminate, migrate and keep the customer job operating?

These fields describe the institutional boundary around a supplier. They are not contract language or legal advice; the final terms remain specific to the buyer, supplier and governing arrangement.

A useful vendor comparison adds one row for each layer it covers and one row for every adjacent responsibility left with the buyer. A polished demo is not a substitute for those boundaries.

For example, a browser provider may supply secure execution and detailed session logs but leave merchant identity, payment authority, result acceptance and refunds to the buyer. The procurement record credits the supplied capability and exposes those four adjacent responsibilities instead of calling the product an end-to-end agent platform.

Where we see the market going

The market will expand horizontally across models, runtimes, tools, protocols, payments and services, while control remains concentrated in a smaller number of institutional systems. Identity directories, merchant systems, enterprise software, payment networks and finance records will continue to decide who can act and whether the result counts.

The largest new categories will form at the joins. Discovery must connect to a current offer. Authority must survive into the effect. Payment must reconcile with delivery. Delegated work must remain attributable across organisations. Companies that make those joins legible will shape the market more than companies that add another isolated agent surface.

About the author and editorial record

Agentic Economy is the accountable publisher of this report. Joel Chan founded the publication in Perth to research the infrastructure, companies and operating choices shaping the agentic economy in Australia. This chapter is source-led market analysis, not a certification, sponsored ranking or claim that every named product has been independently tested.

Previous: AI agent adoption and measurement.

Part II examines the agent-system layers in detail, beginning with models, inference and compute.