Every prior automation wave replaced human muscle with machines. AI agents replace human cognition with software: sourcing suppliers, executing trades, negotiating contracts, managing projects. When the cost of an agent performing a cognitive task drops below the cost of a human performing it, the economic logic that holds corporations together, Ronald Coase's 1937 theory of the firm, inverts. Agents are already in the economy. The open question is what happens to the organizations built for a world where humans were the only economic actors.
Tap to enlargeFrom Tools to Economic Actors
MarketsandMarkets put the AI agents market at $5.26 billion in 2024 and $7.84 billion in 2025, with a $52.62 billion 2030 print at a 46.3% compound rate. That is a paid TAM model, not a census. A PwC survey of 308 US executives in April 2025 found 79% saying AI agents were already being adopted in their companies. That is a US executive sample, not a global production census, and PwC's own write-up notes that adoption is often shallow. Something structural has still changed: software now operates on people's behalf.
AI agents that can perceive, reason, plan, and execute complex tasks autonomously represent something different from prior automation waves. They interpret ambiguous goals, formulate multi-step plans, and adapt strategies based on changing conditions, rather than following a predetermined script.
AI Agent Market Size
Projected market value, 2025-2030 (33-46% CAGR)
Sources: MarketsandMarkets, BCC Research, PixelBrainy (2025-2026 estimates). Some projections reach $250B+ by mid-2030s.
The distinction between automation and agency is structural. A script that scrapes data is a tool. A program that executes a trade at a price threshold is automation. An AI agent tasked with "finding the best inflation hedge" can search financial databases, run macroeconomic simulations, analyze market sentiment, execute trades across platforms, and adjust its portfolio in real time. No individual step requires human intervention. The delegation is of a cognitive function.
Enterprise AI Agent Adoption
Key metrics, 2025-2026
Sources: Gartner (2025-2026), Accelirate enterprise survey. "Apps embedding agents" = Gartner forecast for enterprise applications with task-specific AI by 2026.
The Coasean Inversion
Ronald Coase's 1937 theory of the firm proposed that companies exist because transaction costs in the open market (finding suppliers, negotiating contracts, monitoring quality) exceed the cost of performing those functions internally. When it becomes cheaper to hire employees than to contract externally, firms grow. When the reverse is true, firms shrink.
This logic has been tested before. In the 1850s, a Chicago grain merchant who wanted to buy wheat from a farmer in Kansas faced weeks of delay: sending letters, dispatching agents on horseback, inspecting quality in person, negotiating price without knowing what competitors offered. The telegraph, commercialized in the 1840s and reaching 23,000 miles of wire in the U.S. by 1852, compressed those information costs to minutes. By 1865, the Chicago Board of Trade had formalized standardized futures contracts, replacing face-to-face negotiation with rule-based, protocol-mediated exchange. Firms that had existed primarily to manage information asymmetry (commodity brokers, regional middlemen) shrank or disappeared. New institutions, the futures exchanges themselves, emerged to coordinate economic activity at lower cost.
AI agents repeat this pattern at a different scale. When an agent can discover, vet, negotiate with, and execute contracts with other agents in seconds, the transaction costs that justify large organizations approach a fraction of their current levels. The economic argument for keeping thousands of employees in rigid departmental structures weakens. The telegraph collapsed costs from days to minutes. Agents collapse them from hours to milliseconds, and they do it across cognitive work.
The Coasean Shift
How AI agents reduce the transaction costs that justify the firm
| Function | Legacy cost | With agents | Mechanism |
|---|---|---|---|
| Supplier discovery | High | Near zero | Agents search, vet, compare across markets |
| Contract negotiation | Medium-High | Low | Agent-to-agent negotiation (A2A) |
| Quality monitoring | High | Low | Continuous automated verification |
| Market research | Medium | Near zero | Real-time data synthesis across sources |
| Coordination | High | Low | Multi-agent orchestration frameworks |
Framework: Coase (1937). Application to AI agents: Berkeley Haas (2025), NBER working papers (2025-2026).
When transaction costs collapse, economic activity can shift from centralized hierarchies toward networks of specialized agents that assemble and dissolve as needs change.
This does not mean all firms vanish. A 2025 California Management Review essay from Berkeley Haas argues that multi-agent systems need an orchestration layer, and that whoever runs that layer sets the rules. That can concentrate power even as transaction costs fall. The telegraph killed small-town commodity brokers and concentrated power in the Chicago Board of Trade. Same fork: dispersal, or a new clearinghouse.
What a Firm Could Look Like
A product launch in an agent economy might involve contracting a market analysis agent from one provider, a brand identity agent from another, a swarm of distribution agents from a third, and a supply chain logistics agent from a fourth. These agents coordinate through standardized protocols (MCP, A2A), negotiate resource allocation, and execute strategy. The "firm" becomes a temporary, task-specific nexus of contracted agentic services, dissolving and reforming as needs change.
While AI agents can reduce micro-level transaction costs, their fragmented adoption within a firm can lead to duplicated efforts, conflicting processes, and loss of coherence. Gartner warned in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls. That is a forecast, not a body count. A "guardian agent" layer is vendor talk for oversight software. No primary census puts it at 10-15% of the market by 2030.
The Protocol Layer
No agent economy functions without interoperability. An agent that can only communicate with other agents from the same vendor is a chatbot with extra steps. The critical infrastructure is the protocol layer: shared standards that let agents from different providers discover each other's capabilities, negotiate terms, and exchange structured data.
Two protocols have emerged as leading candidates. Anthropic released the Model Context Protocol (MCP) in late 2024, providing a standardized way for agents to connect to external tools, databases, and services through a common interface. Google followed with Agent-to-Agent (A2A) in April 2025, designed specifically for inter-agent communication: discovery, task delegation, status updates, and result exchange between agents that may run on entirely different models and infrastructure. By early 2026, MCP had attracted integration from over 30 tooling providers, and A2A had backing from Salesforce, SAP, and several enterprise platform vendors.
The historical parallel is HTTP. In 1991, Tim Berners-Lee published the specification for HTTP/0.9 at CERN. Before HTTP, networked computers could exchange files through FTP, Gopher, and proprietary protocols, but each required specialized clients and offered no unified way to link documents across servers. HTTP made networking composable: any document could link to any other document, regardless of server software, operating system, or institutional affiliation. The web exploded not because of any single application but because the protocol layer removed coordination costs between publishers.
MCP and A2A aim to do for agents what HTTP did for documents. An agent built on Claude can delegate a subtask to an agent built on Gemini, which can query a tool exposed through an MCP server, all without any of the three sharing a codebase, a vendor, or even a common model architecture. The composability matters more than any individual agent's capability. A mediocre agent with broad protocol access outperforms a brilliant agent locked in a silo.
HTTP took roughly four years (1991 to 1995) to go from specification to mass adoption, catalyzed by the Mosaic browser. MCP and A2A are on a faster track: within 18 months of release, both had significant enterprise adoption. The difference is that agent protocols do not need a consumer interface to prove value; they prove it in reduced integration costs and faster deployment cycles.
The risk is fragmentation. If MCP and A2A do not converge or at least bridge cleanly, the agent economy could split into incompatible camps, each vendor's agents unable to coordinate with the other's. This happened with instant messaging in the 2000s (AIM, MSN, Yahoo, ICQ, all walled off), and the cost was a decade of lost interoperability until XMPP and later open APIs partially resolved it. The protocol layer is not glamorous, but it determines whether the agent economy becomes a connected network or a collection of proprietary silos.
The Labor Recomposition
The agent economy commoditizes entire cognitive workflows. The work of a junior analyst, a paralegal, a market researcher, or a mid-level project manager can be decomposed into information-gathering, synthesis, and execution steps that agents handle efficiently. The economic pressure to unbundle these roles is real, and the evidence is already accumulating.
Klarna said on 27 February 2024 that its OpenAI-powered assistant handled 2.3 million customer-service conversations in its first month, two-thirds of chats, equivalent work of 700 full-time agents, and an estimated $40 million profit improvement in 2024. That is Klarna's workload math, not a layoff count.
Klarna is not an outlier. Salesforce has claimed high no-human resolution rates for Agentforce in early deployments. Those are vendor case studies. McKinsey's State of AI has repeatedly found most surveyed organizations using AI in at least one function, with customer service, marketing, and software engineering among the top deployments. The exact percentage moves with each survey wave.
The value of human labor concentrates in tasks that resist commoditization: high-level strategic direction, genuine creative ideation, complex ethical judgment, and deep empathetic connection. The premium lands on the ability to ask the right questions, define goals that agents pursue, and provide context that data alone cannot capture.
New roles are forming around the seams. "Prompt engineer" appeared on job boards in volume starting in 2023. LinkedIn job-search counts circulate in the thousands for US listings with that title. Those are search snapshots, not a BLS occupation. Agent orchestrators, people who design multi-agent workflows, set guardrails, and manage inter-agent coordination, have become a recognized function at companies like Cognizant and Accenture. AI auditors, tasked with reviewing agent decisions for bias, compliance, and accuracy, represent a growing compliance need, particularly in finance and healthcare where regulatory scrutiny is high. The labor market is recomposing around the boundary between what agents do well (high-volume, pattern-matching, data-dense tasks) and what humans still own (ambiguity, judgment, trust).
The "true cost" of AI agents can run well above LLM token costs. Enterprise deployments face substantial expenses in data preparation, legacy system integration, security, compliance, and ongoing maintenance. This total cost of ownership determines whether firms should build or buy agentic capabilities.
From Attention to Intention
For three decades, the internet has been built around capturing human attention. Interfaces, content, and business models all chase eyeballs. The agent economy operates on a different axis. Agents have objectives. They move through APIs, structured data, and semantic metadata to fulfill goals with minimal friction.
This creates pressure to rebuild digital infrastructure. A hotel booking service's value shifts from its photo gallery to whether a travel-planning agent can query availability, pricing, and attributes through a single API call. A news service's relevance depends on providing semantically tagged, machine-readable content that a research agent can ingest, not on clickbait headlines.
A web built for agents is one where structured data outweighs unstructured prose, cryptographic proof outweighs unsubstantiated claims, and efficiency outweighs engagement.
This shift could produce a more honest digital commons. Agents, being logical and goal-driven, are less susceptible to dark patterns, misinformation, and sensationalism. They favor reliable sources, verifiable data, and transparent services. Economic incentives realign from capturing human attention to servicing machine intention.
The Risks
Concentration. The companies that control foundation models and orchestration platforms can become the access point for economic activity. Whoever owns that layer owns the gate.
Systemic fragility. A cascading failure in interconnected financial agents could trigger a market crash in seconds. The speed of agent-to-agent interaction means errors compound exponentially before humans can intervene.
Accountability gaps. When an autonomous agent causes harm, is the user, the deploying company, the model trainer, or the data provider responsible? Existing legal frameworks, built around human actors and corporate entities, do not have clear answers. New rules of algorithmic accountability are needed.
Malicious agents. Agents designed for scams, cyberattacks, or large-scale manipulation could create hostile environments that undermine the efficiency the agent economy promises. Agent verification (KYA, or "Know Your Agent") is emerging as a parallel to KYC.
A hotel's value to an agent is an API that returns availability and price, not a photo gallery built to hold a human's eye. Gartner forecast in August 2025 that 40% of enterprise apps would have task-specific agents by the end of 2026. That is a forecast, not a shipped census. The open risks are concentration at the model layer, cascades faster than a human can halt, and no settled rule for who is liable when an agent causes harm. KYA (Know Your Agent) is the early name for that last gap. Protocol convergence and that liability rule decide whether this stays a market or becomes a few platforms with a crash mode measured in seconds.