What is the agentic economy?

On this page

Chapter 1 of The Hitchhiker's Guide to the Agentic Economy: Australia 2026

Report hub · In this chapter: category tiers · delegated-work model · worked offer · bounded job

The delegation problem

Every commercial institution we take for granted was built to solve a single, ancient problem: delegation.

When a merchant in renaissance Venice sent a supercargo on a trading voyage to Alexandria, or when a modern CFO authorizes a procurement team to negotiate supplier terms, the underlying challenge is identical. How do you grant someone else the authority to act, buy, and commit resources on your behalf, while ensuring they stay within bounds, leave an honest record, and answer for the outcome when something fails?

Over centuries, trade developed a dense web of machinery to manage that principal-agent relationship. Contracts define the boundary of what was promised. Double-entry bookkeeping prevents silent financial leakage. Credit requires assets at risk. Invoices require an authorized buyer to inspect and accept goods before money changes hands. And when an agreement breaks down, courts and arbitration panels enforce liability against identifiable people and corporate entities.

None of those problems are new. What is new is the actor standing at the point of decision.

For most of computing history, software was merely a tool inside that human machinery. Spreadsheets calculated figures, databases stored ledgers, and automated scripts moved records on fixed schedules. Even generative AI primarily functioned as a conversational assistant: an employee prompted a model, reviewed the draft, and made the consequential choice to send, sign, or buy.

Now, software is being granted the authority to make those decisions directly. An agent workflow can evaluate competing suppliers, negotiate an order, issue a payment credential, coordinate external APIs, and update internal enterprise systems without a person approving every discrete action along the way.

The immediate temptation is to treat this purely as an engineering challenge: benchmark reasoning scores, compare context windows, and track protocol drafts. But the friction organizations encounter when moving agents into production is almost never a lack of model capability. It is the collision between an autonomous digital actor and institutional infrastructure engineered entirely for human trade:

• Identity and Authority: Authentication proves which computer connected, but it does not tell a receiving system who authorized the agent to spend money, what budget caps apply, or how to revoke that authority mid-execution if the workflow goes off course.
• Transactions vs. Relationships: An API call can settle a request in milliseconds, but enterprise trade is an ongoing relationship. It requires vendor pre-qualification, tax-compliant invoicing, data privacy boundaries, and an enforceable path for refunds when delivered data is corrupt.
• Activity vs. Acceptance: In software engineering, a 200 OK or a completed execution loop looks like success. In business, activity is just cost until a principal inspects the output and accepts the work.

Software has entered the commercial realm as an economic actor, but it has arrived in an economy whose rules, ledgers, and legal protections assume a person is always in the loop.

Key Takeaways

• Work becomes agentic when software selects among permitted actions and changes a tool, service or business process.
• The practical unit is objective + context + authority + action + evidence + accountability.
• Autonomy is a property of a particular task and boundary, not a permanent label for a product or company.
• A fixed workflow and a model-directed agent can both create value. The right design is the least complex system that can perform the job dependably.
• An accepted outcome is a useful operational endpoint; value and margin still require a baseline, costs and repeat demand.

The category is forming around delegated work

The market is moving beyond assistants that produce answers toward systems that advance work. Model providers are adding tool use and computer control. Software platforms are embedding agents inside existing applications. Framework companies are building orchestration and memory layers. Service providers are beginning to package completed tasks rather than access to software alone.

OpenAI states the category in operational terms: agents are “systems that independently accomplish tasks on behalf of users” (OpenAI, "New tools for building agents"). The key word is tasks. The market is moving from generated text toward completed work, even though providers draw the autonomy boundary in different places.

The common unit underneath those products is delegated work: a defined objective, the context and authority available to the system, the actions it takes, the evidence it leaves and the person or organisation that remains accountable. Products implement that unit differently. It gives the market a clearer boundary than the word agent by itself.

From an answer to an action

Software has acted without a person at the keyboard for decades. Scheduled jobs, rules engines, webhooks, and robotic process automation all update systems automatically. The newer distinction is not action versus passivity. It is whether a model or learned policy can select among permitted actions while pursuing an objective.

Generative AI made the interface conversational. An assistant summarises a contract, drafts a campaign or suggests code while a person chooses the consequential next step.

Agentic systems move that boundary. The system chooses a next step, selects a tool, inspects the result and continues until it reaches a stopping condition. An accounts-receivable agent identifies an overdue invoice, checks the account policy, drafts an appropriate message, records the contact and schedules a follow-up. A purchasing agent compares approved suppliers, requests a quote and prepares an order for human approval. A research agent commissions a bounded data service, validates the returned file and adds it to a briefing.

These systems differ in complexity, but each advances a real process. The business is delegating some combination of selection, sequencing and execution.

Anthropic distinguishes workflows, where code predetermines the path, from agents, where a model dynamically directs its own process and tool use. Its engineering guidance also recommends using simple, composable patterns when they are sufficient (Anthropic, "Building effective agents"). That distinction is useful for design without turning the market into a contest over which product deserves the word agent. A fixed workflow with a clear control surface may be the stronger commercial system.

The Delegated Work Model

A reliable commercial deployment follows a single unit of delegated work across six fields:

objective + context + authority + action + evidence + accountability

Objective

The objective states the job and its stopping condition. "Help with sales" is too broad. "Prepare a sourced briefing on this prospect before tomorrow's meeting" gives the system a deliverable, subject and deadline.

Context

Context supplies the information the system needs: the customer's request, internal records, current policies, product data, prior actions and relevant external sources. Context needs a freshness and provenance boundary. An agent acting on an obsolete price list or an untrusted instruction can execute the wrong decision perfectly.

Authority

Authority defines what the system may see, decide, change and spend. It links the action to a principal, such as a person or organisation, and constrains the agent by role, resource, amount, supplier, time or approval condition. Authentication proves which credential is present. Authority answers the operational question: is this credential allowed to take this specific action now?

Action

The action changes something outside the model. It may call an API, use a browser, send a message, create a record, run code, request another service or initiate a transaction. Each action creates side effects and therefore needs controls for retries, duplication and failure.

Evidence

Evidence records what the system received, decided, called and produced. Useful evidence is more than a chat transcript. It connects the objective, source data, policy, tool response, delivered result, approval and exception so the business can reconstruct the job.

Accountability

Accountability establishes who owns the outcome and how a problem is corrected. A person, team, supplier or institution still carries the obligation to monitor, accept, reject, refund, repair or explain the result. Delegation changes how work is performed; it does not eliminate commercial and legal responsibility.

For consequential work, operational governance requires explicit roles:

Role	Decision owned
Business owner	Job purpose, benefit, customer promise and operating result
Control owner	Authority, data, payment, evaluation or recovery control
Independent challenger	Evidence review outside the delivery team where consequence requires it
Risk-acceptance authority	Residual risk, exception duration and conditions for continued use
Incident owner	Containment, notification, correction, restoration and customer remedy

The model joins technology to economics. A model provider supplies capability. A runtime manages state. A tool or service performs an action. An identity or policy layer limits authority. A payment system transfers value. An enterprise system records the commercial event. A human or organisation remains responsible for the customer outcome.

A practical vocabulary

The market uses overlapping terms. The following distinctions make products and claims comparable:

Term	Working meaning in this report	The decision boundary
AI assistant	Produces information or a recommendation for a person	A person chooses and performs the next consequential action.
Automation	Executes predetermined rules or steps	Code fixes most of the path in advance.
AI workflow	Combines models and deterministic steps in a defined process	The workflow may branch, but the system's range is designed explicitly.
AI agent	Chooses and carries out some steps towards an objective using permitted tools and information	The model can select among actions inside a boundary.
Multi-agent system	Uses several specialised agents or agent roles to divide and coordinate work	Coordination and hand-off state become part of the system.
Agentic commerce	An agentic system crosses a buyer-supplier or payment boundary during discovery, ordering, delivery or support	The report names which commercial states and authority the system actually handles.
Agentic economy	The broader system of production, coordination and exchange that forms as software takes on bounded work	Includes infrastructure, firms and institutions around the action, not just the agent product.

None of these forms is automatically superior. A deterministic automation can outperform a flexible agent on a stable, high-volume task. A human-approved assistant remains the appropriate design where judgement, regulation, or liability requires human sign-off. The relevant questions are whether the system performs the intended job, how often it fails, what intervention it needs and whether the economics improve.

Autonomy belongs to the task

Calling a product autonomous obscures the design choices that matter. The same system may work unattended for one task, require approval for another and have no authority at all in a third.

Autonomy varies across five operating dimensions:

Dimension	Lower delegation	Higher delegation
Planning	Person supplies each step	System selects and revises the path
Tool choice	Fixed tool and parameters	System selects tools and parameters
External action	Draft or recommendation only	System changes records or communicates externally
Spending	No payment authority	Bounded or policy-controlled spend
Review	Every action approved	Exceptions or samples reviewed

Anthropic's research on autonomy in deployed systems indicates that observed behaviour depends heavily on product design, task structure and human oversight, and that product telemetry rarely captures downstream review (Anthropic, "Measuring AI agent autonomy in practice"). The operational reality is clear: measure the specific task, permitted action set, and review criteria rather than inferring autonomy from a vendor marketing label.

Longer operation also compounds risk. A system executing a single read-only lookup has a negligible failure surface. A system running for hours, invoking external tools, altering databases, and spending money accumulates state and consequence. Reliability at each step compounds across the path. The organisation needs checkpoints, budgets, timeouts, idempotency keys, escalation paths, and recovery mechanisms to resume or reverse work.

The agent is a system, not a model

A capable model is necessary for many agentic tasks, but the operating system around it determines whether the work survives contact with an actual business:

Layer	Job in the system	Typical failure if it is missing
Model and inference	Interpret objectives, reason and choose steps	The system cannot handle ambiguity or adapt its plan.
Context, memory and data	Supply current information and retain relevant state	The system acts on missing, stale or unauthorised information.
Runtime and orchestration	Manage execution, state, retries, hand-offs and long-running work	The process stalls, repeats or loses state.
Tools and service access	Perform actions in software and the world	The agent can recommend but cannot complete the job.
Identity, authority and policy	Establish the principal and constrain permitted action	The system overreaches or cannot prove permission.
Evaluation and observability	Test quality and record what happened	Failures remain invisible or cannot be explained.
Security and recovery	Contain hostile input, data loss and operational failure	A small error becomes a business incident.

An agentic AI harness is the surrounding software environment that gives a model tools, memory, orchestration and controls. The Australian Signals Directorate uses this term in guidance focused on risks around the model, including tool permissions, untrusted content, data access and monitoring (ASD, "Agentic AI harnesses"). This systems perspective reflects commercial reality: a customer buys a dependable business result, not a raw neural network in isolation.

Where the economic change appears

Delegated action alters four core business surfaces:

The interface becomes an objective

Users previously had to navigate software menus, forms, and workflows. An agent can receive a high-level objective and operate multiple backend systems on the user's behalf. The user interface shifts from "which button do I press?" to "what outcome do I require, and what authority should the system receive?"

The product can become a completed task

Software businesses have traditionally monetized through seat licenses or consumption quotas. When software performs more of the operational work, suppliers can package an outcome: a reconciled account, a qualified lead list, a checked contract, or a resolved support request. The commercial unit shifts toward completed work, although outcome pricing is viable only when performance standards and liability for failure are unambiguously defined.

Distribution becomes machine-mediated

Agents discover tools, products, and services through registries, protocols, assistant ecosystems, marketplaces, and direct APIs. A supplier's offer must be machine-legible: who it is for, what inputs it accepts, what it returns, what it costs, and what recourse exists if delivery fails. Being reachable on a network does not guarantee transaction volume; ranking, proven reliability, and verifiable credentials determine whether a supplier makes the shortlist.

Management becomes authority design

Managers already decide who can approve an expense, contact a customer, or alter a production record. Agentic systems make those governance decisions executable in software policies. The work requires defining budgets, permission boundaries, review thresholds, auditable logging, and escalation owners. This is organizational operating-model design translated into software architecture.

Reference job AE-JOB-001: a competitor briefing

Consider a Perth advisory firm that repeatedly prepares competitor briefings for corporate clients. This synthetic reference job is AE-JOB-001, and it serves as the concrete benchmark across every chapter of this report. The firm wants to make its research capability programmatically accessible to client workflows while retaining mandatory professional review:

Element	Design
Objective	Produce a sourced briefing on a named company and market by a stated deadline.
Context	Client request, approved source classes, geography, prior briefings and confidentiality rules.
Authority	Access only to the submitted client material; spend capped for external data; publication prohibited.
Action	Search approved sources, buy a bounded data input if needed, draft the briefing and route it to an analyst.
Evidence	Source links, retrieval dates, data-service receipt, change log and analyst decision.
Accountability	The advisory firm owns accuracy, delivery and correction; the analyst accepts or rejects the draft.

The firm packages that operational process into a formal commercial offer:

Offer field	Commercial term
Job	One sourced competitor briefing using up to ten approved public sources
Price	A$750 including GST
Delivery	One business day after a complete brief is accepted
Review	Analyst checks sources, unsupported claims and the final recommendation
Variable spend	External research purchases capped at A$25 unless the buyer approves more
Acceptance	Required sections present; every material factual claim sourced; no prohibited client data used
Remedy	Rejected work corrected within two business days or the fee refunded
Data boundary	Client material retained only for the agreed delivery and correction period

These terms isolate the operational economics hidden behind the phrase AI service. If fulfilling the job consumes A$18 in model and data fees plus 45 minutes of analyst review, the firm generates a viable contribution margin. If analyst review swells to three hours because the draft is hallucinatory or unstructured, the firm has an operational loss regardless of model speed.

The agentic component is not a conversational chat interface. It is the controlled progression of a job through automated research, service purchasing, drafting, and human sign-off. The commercial deliverable remains a familiar professional briefing with a clear service-level agreement; software makes the intake frictionless and fulfillment repeatable.

This reference job also illustrates why the entire stack matters. Delivering AE-JOB-001 requires a foundation model, specialized search APIs, an orchestration engine, a payment credential, an accounting integration, and an observability pipeline. Each vendor supplies one component; the client purchases a single completed deliverable.

NIST describes the prevailing agent pattern as a foundation model wrapped in software scaffolding that enables it to operate external tools beyond raw text generation (NIST, "Tool Use in Agent Systems"). NIST emphasizes that tool capabilities and constraints must be formally communicated across the supply chain, reinforcing the necessity of evaluating the entire operating system rather than model intelligence alone.

What counts as completion

The lowest bar for agentic software is generating an unverified response. In production commerce, a completed job requires an auditable progression:

1. The system receives a validated objective.
2. It operates within explicit context boundaries and authorized limits.
3. The required actions execute idempotently without unhandled errors.
4. The deliverable is transmitted to the receiving system.
5. The buyer or authorized verification system formally inspects and accepts the work.
6. Financial, tax, and operational ledgers reconcile.
7. The customer repeats or expands usage.

Distinguishing these stages is essential because the easiest metrics to capture in software are intermediate signals. A tool invocation merely logs execution. A payment confirms an asset transfer. An HTTP response confirms network delivery. Only formal acceptance confirms that the customer received what was contracted. Long-term commercial viability requires tracking the complete ledger: input costs, exception rates, rework overhead, and repeat transaction volume.

The bounded job is the market's useful unit

Eleven operational fields define a delegated job, distinguishing a reliable commercial workflow from a fragile tech demo:

Field	What it establishes
Delegated job	The exact result the system is expected to produce
Economic actors	Who requests, pays for and accepts the work
Discretion	Which choices the software can make
Reach	Which data, systems and spend it can access
Completion evidence	What proves delivery and acceptance
Exception ownership	Who reviews and remedies failure
Cost base	Which costs determine whether the job is viable
Baseline	Current volume, labour time, error rate and full cost
Investment	Implementation, security and change cost to recover
Return boundary	Benefit, maximum loss and payback rule
Decision rights	Who owns funding, expansion, redesign and termination

Different interface protocols solve distinct boundaries across this lifecycle:

Requirement	Standard mechanism
Execute a single deterministic operation	REST / gRPC API
Connect an agent to local tools, data, or context	Model Context Protocol (MCP)
Delegate a multi-step task across independent agents	Agent-to-Agent (A2A) / task protocol
Manage product catalog, cart, and order states	Commerce protocol (e.g., UCP, ACP) or merchant API
Authorize and settle transactions	Payment rails aligned to value, velocity, and dispute recourse
Maintain institutional judgment on high-consequence work	Human-in-the-loop review workflow

Software deployments will ultimately be measured against the existing business process they replace rather than synthetic LLM benchmarks. Acceptance rates, human intervention hours, total operational cost, and failure severity determine whether autonomous delegation expands or contracts within an enterprise.

The market trajectory

The generic label agent will increasingly give way to distinct, specialized software layers: conversational assistants that prepare briefs, workflow orchestrators that route multi-step processes, automated workers that execute narrow tasks, and managed services that deliver guaranteed business outcomes.

This shifts commercial advantage toward the operational systems surrounding the model. Providers that can enforce authority boundaries, provide immutable audit trails, and guarantee accepted delivery will capture enterprise spend. Vendors that offer only conversational interfaces without operational safeguards will remain commoditized software features.

Chapter 2 evaluates the empirical measurement problem: distinguishing speculative headline demand from verifiable production adoption.

Sources and method

• Primary sources: Technical specifications and operational guidance from OpenAI, Anthropic, the Australian Signals Directorate, and NIST, evaluated against observable enterprise commercial states. Sources checked as of 13 September 2026.
• Analytical models: The Delegated Work Model, the eleven-field job framework, and the AE-JOB-001 reference job are original analytical structures developed to evaluate commercial trade-offs across market layers.
• Evidence boundary: Provider documentation establishes vendor-stated system behavior; independent operational adoption and reliability are evaluated in Chapter 2.
• Editorial record: Published by Agentic Economy, Perth, Western Australia. Updates and corrections are logged via the editorial record.

Next: Demand, adoption and the measurement problem.