The Hitchhiker's Guide to the Agentic Economy

On this page

Global infrastructure edition, 2026

Agentic Economy's map of the companies, infrastructure and commercial systems shaping delegated software work.

This guide explains how we see the market: the players, what they have built, the categories forming around them, the approaches that matter and the gaps that remain unresolved.

The report's central finding: the agentic economy forms around delegated work, but value appears only when capability, authority, execution, commerce and accountability remain joined to an accepted result.

Parts I–III: market, infrastructure and commerce · Published by Agentic Economy · Research cut-off: 13 September 2026

Imagine a Perth advisory firm that sells a competitor briefing. The job has a named buyer, five named competitors, up to ten approved public sources, a one-business-day delivery promise, an analyst review and a correction policy. Software can research and draft the brief, buy a bounded data input and route the result for approval. The customer still buys a familiar professional service. What changes is how the work is ordered, performed and evidenced.

Agentic economy is the market that forms as software carries work forward under delegated authority. It includes the agents themselves and the models, tools, data, service providers, payment systems, enterprise software, people and institutions required to make their actions useful and accountable.

The important change is easy to miss. Generative AI produces an answer, image or block of code. An agentic system can also choose among permitted steps and act: update a record, call a service, use a browser, coordinate another system, place an order or advance a business process.

That changes the market around software. Models still matter, but agents are assembled systems. They need information and memory; tools and permissions; testing and security; and commercial records and people who remain responsible for the result.

This edition contains Parts I–III: the market frame; seven infrastructure chapters that follow one delegated job from model and context through runtime, service access, coordination, authority and control evidence; and five exchange-and-commerce chapters that carry the same job through discovery, order, payment, enterprise relationship and service economics. The infrastructure and vendors are global; the operating lens is Australian where it changes the market. The company atlas, casebook and Australian sector studies remain future editions rather than implied findings in this one.

Agentic Economy's view: a system is not production-ready until the job, authority, delivered result, accepting person or system and failure remedy can all be named.

Key Takeaways

The practical unit is delegated work: an objective, context, authority, action, evidence and accountability.

An agent is a system, not a model. Its quality depends on the full path from the objective to an accepted outcome.

Availability is ahead of measured adoption. Specifications, product launches and working endpoints are plentiful; independent evidence of repeat customers, accepted work and viable margins is scarce.

Payment is one event in commerce. Enterprise use also needs identity, authority, terms, delivery, acceptance, invoicing, reconciliation and recourse.

The market will use several interfaces and rails. APIs, MCP, A2A, commerce protocols, cards, bank payments and stablecoins solve different boundaries.

The market is moving first through bounded workflows with observable business results, not broad autonomy programmes.

An open observatory looking into a field of stars, representing a guide built to scan a large emerging market

Sources and corrections

Agentic Economy built this guide from primary specifications, provider documentation, official Australian sources and named external research. Every material record carries a checked date and a limit on what the evidence proves. The report distinguishes documented supply from observed use and accepted business outcomes. About Agentic Economy explains the publication and its Australian focus. Readers can submit a source correction through the contact page.

Six findings from Part I

1. Agentic work is a distinct operating mode, not a synonym for automation

Software has executed rules without a person at the keyboard for decades. The newer boundary is model-directed choice: an agent can select among permitted actions while pursuing an objective. Agentic Economy therefore separates assistants, deterministic automation, model-assisted workflows and agentic systems rather than placing them in one inflated category. Chapter 1 defines the tiers.

2. Agents are assembled systems

A model can propose an action. An operational agent also needs current information, retained state, permitted tools, identity, policy, evaluation, records and a recovery path. Australia's cyber-security guidance similarly treats the agent harness around a model as the relevant control surface (Australian Signals Directorate, "Agentic AI harnesses").

3. Capability evidence is ahead of economic evidence

Agentic Economy's source review found extensive specifications, product documentation and working interfaces. Public evidence becomes thinner at accepted delivery, repeat use and contribution margin. This is a finding about the reviewed source set, not a market-size estimate. Chapter 2 shows how the evidence differs.

4. Payment is one event in a longer commercial relationship

ACP and UCP describe commerce states. AP2 describes mandate and payment-authority semantics. x402 and MPP initiate or settle machine payments. Contracts, delivery, acceptance, tax, invoicing, reconciliation and recourse remain separate states. For example, a one-off digital call and a recurring enterprise supplier relationship need different operating models (Stripe Link CLI; Stripe usage-based billing; Stripe Invoicing).

5. Australia is early, active and poorly measured

The Reserve Bank of Australia's liaison received 105 responses from medium and large firms between June and August 2025. It found mostly piecemeal AI use, low practical agentic adoption and mixed realised returns. This is useful business evidence, but it is not a representative census of Australian firms or small service businesses (RBA, "Technology investment and AI").

6. Accepted work is the most useful operating endpoint

Agentic Economy proposes accepted jobs as the bridge between technical and commercial measurement. An accepted job is work that a responsible person or system has judged complete against the promised result. It still needs to be read beside cost, human effort, incidents, rework and repeat use. Chapter 2 provides the job-level record.

The whole market in one view

An agent system is the model plus the context, runtime, tools, authority, evaluation and recovery controls required to carry work forward. The agentic economy can be read as four connected systems.

System	What it contains	The question it answers
Work and demand	Tasks, workflows, sectors, buyers, users, outcomes and repeat use	What useful job is being delegated, and for whom?
Agent systems	Models, compute, context, memory, runtimes, tools, identity, policy, evaluation and security	What lets software decide and act within a boundary?
Exchange and commerce	Discovery, offers, marketplaces, checkout, payments, contracts, billing, delivery and recourse	How do buyers and suppliers form and complete an economic relationship?
Institutions and consequences	Enterprise systems, firms, labour, capital, standards, law and public infrastructure	Who remains accountable, where does value move and what changes around the system?

Across those systems, five things move:

Work: the objective, task state and delivered result.
Data: context, instructions, records and evidence.
Authority: identity, permission, policy, approval and revocation.
Money: quote, credential, payment, invoice, settlement and refund.
Accountability: logs, acceptance, exception handling, audit and recourse.

The market works only when those flows remain joined. A tool call without authority is unsafe. A payment without a deliverable is incomplete. An output without acceptance is not a finished customer job.

What the sources establish

The report separates product capability, market activity and measured outcomes. A detailed specification establishes what a system is designed to do. It does not establish adoption. A large first-party sample shows activity inside one provider's boundary. It does not automatically describe the wider market.

Dimension	Question	Examples
Observed event	What happened?	Specified, available, invoked, paid, delivered, accepted, reconciled, repeated
Source origin	Who produced the record?	Provider, customer, Agentic Economy test, third-party researcher, official body
Study design	How was the claim produced?	Product telemetry, reproducible test, survey, interview, case record, census

Quantitative market claims identify their date, population, counted event and denominator. Product and protocol claims identify the version or checked date. Agentic Economy analysis appears as analysis rather than external consensus.

Provider documentation and specifications establish documented supply. They do not, by themselves, establish adoption, interoperability, service quality, control effectiveness or business value. The contracts and records introduced in Part II are Agentic Economy proposals for making decisions visible; they are not standards, certifications or substitutes for jurisdiction-specific legal and assurance work.

Read the report

Part I: The market

Part II: Agent systems and infrastructure

Part II maps the documented global supply of agent infrastructure and shows how its layers connect. Its provider tables record specifications, supported capabilities, commercial surfaces and known limits at the checked date. The tables describe the market; they do not rank providers or convert product availability into proof of adoption, reliability or business value.

The chapters follow one job through the system:

model and context
  -> controlled run
  -> external capability
  -> optional agent delegation
  -> authority enforced at the effect boundary
  -> joined evidence and acceptance

The opening competitor-briefing job is the running illustration, not observed evidence. Chapter 6 asks whether its research and drafting run can pause and recover. Chapter 7 constrains approved sources, any paid data input and the CRM route. Chapter 8 applies only if a separately operated specialist agent receives part of the research. Chapter 9 binds each source, spend or record change to current authority. Chapter 10 joins the trace, resource receipts, analyst review, correction effort, cost and final customer acceptance. A failure at one layer stays attached to the same job rather than becoming a separate technical success.

Chapters 6, 7, 9 and 10 form the core production path: runtime, service access, authority and control evidence. Chapter 8 becomes material when work crosses into another independently operated agent. The control artefacts are linked views of one job, not five competing standards. Chapter 10 defines their ownership and join keys.

A supervised, read-only and reversible draft has a smaller control surface than a system that writes records, runs unattended or delegates work remotely. The required infrastructure expands with the consequences the job can create.

How infrastructure enters production

Stage	Typical boundary	Control surface that appears next
1. Supervised read	One read-only direct API, approved context and a person accepting the result	Correct source and account, denied cross-boundary read, trace, cost and acceptance record
2. Approved effect	One narrow write with a shown effect, current authority, idempotency and resource receipt	Invalid-input, approval-change, timeout, duplicate and revocation tests; no unresolved write
3. Unattended or long-running work	Durable run state, retry budget, cancellation, reconciliation owner and incident path	Worker-loss, dependency-outage and recovery tests; named stop authority and retained evidence
4. Remote delegation	One approved agent boundary with no implicit subdelegation	Capability freshness, protocol conformance, scoped data, child-task lineage, acceptance and recourse

This is a common expansion pattern, not a universal sequence. High-consequence work can require the full control set before its first live action. Reversible, low-risk drafting may never require remote coordination.

Each expansion also adds commercial requirements:

Commercial field	Decision required
Total cost	Expected cost per accepted job, including integration, model and tool use, storage, review, recovery and change maintenance
Operating cover	Named service owner, approval cover, reconciliation queue, incident response and any after-hours requirement
Supplier support	Support window, response target, change notice, escalation path and responsibility for a failed or duplicated effect
Data and evidence	Retention, region, protected access, export format, deletion and availability during a dispute or provider exit
Acceptance and remedy	Who accepts or rejects work, what rework is included, when billing occurs and which correction, refund, restoration or liability path applies
Exit	Credential removal, data export, open-job treatment, replacement route and the cost of migration

The manuscript does not supply universal price or staffing ranges. Those depend on job volume, consequence, provider terms and the operating model. Where evidence does not exist, an empty field is more accurate than a manufactured industry average.

Systems that affect customers, workers, citizens or other people also sit inside jurisdictional and public-interest constraints. The following frame is Agentic Economy's view of the recurring issues, not a universal legal test.

Public-interest field	Decision required
Affected people and purpose	Who may be affected, what lawful and legitimate purpose is served, and who is excluded or disadvantaged?
Data and jurisdiction	Which privacy, data-location, cross-border, employment, consumer, accessibility and sector rules apply, and which organisation is responsible for each?
Notice and contest	What must be disclosed, and how can an affected person understand, challenge, correct or appeal the result?
Incident and regulator	Who can stop the job, preserve evidence, notify affected parties and contact the relevant regulator or authority?
Remedy	Who must correct the record, restore the prior state, redo the work, refund or otherwise answer for harm across supplier boundaries?

Buyer acceptance does not settle legitimacy for everyone affected. The applicable rules come from the places, sectors and relationships in which the job operates; this report is not legal advice.

Part III: Exchange and commerce

Part III maps the systems forming around exchange and commerce. It follows one synthetic first job, AE-JOB-001, through the commercial chain so that discovery, authority, payment, delivery and service economics remain attached to the same unit of work. Each chapter adds one state to the job rather than introducing a separate system:

find a candidate -> verify buyer, seller and offer -> authorise and order
-> pay and reconcile -> deliver and accept -> repeat, amend or exit

Chapter	New state and decision	Authoritative owner
11. Discovery and offers	Candidate becomes a current, bounded quote—or stays unknown	Discovery source, buyer and supplier
12. Commerce protocols	Approved offer becomes a merchant order without duplicate or silent reprice	Buyer policy and merchant order systems
13. Payment and settlement	One authorised payment is initiated, resolved and joined to the job	Processor, bank, wallet/network and finance
14. Enterprise relationship	Job sits inside qualification, contract, support, invoice, remedy and exit	Procurement, contract, service, AP and data owners
15. Service economics	Metered activity becomes delivered and accepted work with measurable cost and repeat evidence	Supplier operations, reviewer and finance

The Commercial Job Record is the canonical join, not a replacement ERP: it stores stable references to authoritative native records. Outcome unknown means evidence is missing and a named owner must reconcile it before a retry or closure. MoR means merchant of record; PSP means payment service provider; MCP is Model Context Protocol; A2A is Agent2Agent; UCP, ACP and AP2 are commerce or payment-authority protocols; x402 and MPP are machine-payment protocols. The chapters define each term where it changes the decision.

Part III follows the competitor-briefing job from a visible supplier to a current offer, versioned order, bounded payment route, accepted delivery, reconciled relationship and repeat-use decision. Its five artefacts are linked views of one Commercial Job Record. They are editorial proposals, not competing standards or substitutes for commercial, legal, accounting or assurance records.

Planned Part IV: Applications, firms and institutions

Use cases; enterprise software; organisational change; labour and skills; standards, law and public infrastructure; capital, competition and value capture.

Planned Part V: Decisions and forward view

Role-specific decisions; signals and falsifiable bets; methods, limitations and corrections.

The Australian lens

Australia is neither a separate technical stack nor a footnote to a US market. Australian businesses use global models, clouds, agent frameworks and commerce platforms. Their deployment choices also sit inside local privacy obligations, financial regulation, employment conditions, payment infrastructure and industry practices.

Every chapter therefore asks four local questions:

Is the product available to Australian buyers or builders?
Which Australian companies, customers, institutions or deployments are visible?
Which local systems or rules alter the path to production?
Which Australian evidence measures real use and outcomes, and which gaps remain?

Current evidence supports a measured starting point. Agentic capability is arriving quickly. Broad, independently measured economic adoption is not. Australian activity is forming around useful workflows, explicit authority, observable outcomes and controlled failure paths while the wider market takes shape.

One global survey shows why activity and outcomes must stay separate. In McKinsey's 2026 sample, 40 per cent of respondents at large organisations said their organisations were scaling AI agents, while 37 per cent of all respondents reported any positive EBIT impact from AI (McKinsey, "The State of AI in 2026"). Those figures use different denominators and do not measure the accepted-job rate defined in this report.

How Part I maps the market

Market question	Chapter	Agentic Economy framework
What makes the market agentic?	What is the agentic economy?	The Delegated Work Model and bounded-job unit
What counts as adoption?	AI agent adoption and measurement	Job-level measures that separate activity from accepted economic outcomes
How do the categories connect?	The agentic economy market map	Four systems, five flows and the boundaries between suppliers

For connected introductions already on aecon.ai, see AI agents versus automation, what the agentic economy is and evaluation, observability and security.

About this edition

This is a living field guide and the opening edition of a larger market-report programme. It is not an official statistic, certification, paid ranking or complete company census. Parts I–III combine primary specifications, provider documentation, official Australian sources and named external research. Agentic Economy analysis is labelled. Future editions will extend the company and product coverage, add original fieldwork and deepen the Australian market record without changing the limits of the evidence presented here.

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 edition is source-led market analysis, not a sponsored ranking, legal advice or a claim that every listed product has been independently tested.