In 1937, Ronald Coase asked a deceptively simple question: why do firms exist at all? His answer—that companies internalise work whenever the cost of doing so is cheaper than contracting it out—shaped a century of management theory. It was a framework never designed to accommodate a software agent that can negotiate a supplier contract, draft the resulting document, and file it before a human being has finished their morning coffee. And yet, in the first quarter of 2026, that is precisely the situation.
Two product releases crystallised what technologists had predicted and markets were beginning to price. Google Antigravity—an agent-first IDE powered by Gemini 3—and Anthropic's Claude Cowork, billed with characteristic understatement as 'Claude Code for the rest of your work,' mark the closing of a gap that defined AI's commercial limitations for a decade: the distance between what a model can reason about and what it can actually do. That distance, it turns out, was also the entire economic moat of the global SaaS industry.
The Oracle Becomes an Operative
Earlier AI models were, in the precise technical sense, stateless: each prompt arrived tabula rasa, context forgotten, execution impossible. They were oracles—consulted, then dismissed. The systems reaching enterprise deployment in 2026 are different in kind, not degree. They retain persistent memory, decompose ambiguous goals into executable sub-tasks, invoke tools, evaluate results, and adjust strategy mid-flight. Stateless systems resolve 54% of issues on first contact; stateful agents resolve 89%.
Google Antigravity makes this concrete. Its defining innovation is not a single feature but an architectural philosophy: the IDE as autonomous workspace. Three surfaces—editor, terminal, and a headless browser controlled by a dedicated sub-agent with multimodal vision—allow it to read code it was not asked to inspect, install its own dependencies, and verify outputs against design specifications by literally looking at a running web application. Feature generation tasks that take competitors 68 seconds are completed in 42. Bug introduction rates have halved.
The trajectory, more than any single score, is what should concentrate minds. GPT-4 resolved 33% of benchmark tasks in June 2024. Claude Sonnet 5 ('Fennec') resolved 82.1% in February 2026—a span of twenty months. The inflection has not slowed. It has accelerated.
The SaaSpocalypse: Coase Takes the Stand
Between January 15th and February 14th, 2026, approximately two trillion dollars in market capitalisation evaporated from the software sector. Analysts coined 'SaaSpocalypse' to describe it. The mechanism is Coasean in its clarity.
The SaaS revenue model rests on a simple equation: more employees equals more seats equals more revenue. When a single agent handles the work of five CRM operators, that equation collapses. Research suggests AI agents reduce transaction costs—search, negotiation, enforcement, coordination, monitoring—by 70 to 85 percent. When those costs approach zero, the specialised tools built to manage them lose their reason to exist.
The cybersecurity sell-off deserves special attention for what it reveals about markets under uncertainty. When Anthropic launched Claude Code Security—an LLM-powered codebase scanner that identified more than 500 previously undetected vulnerabilities in open-source codebases—the instinct was to sell CrowdStrike, Okta, and Cloudflare. Barclays called the move 'incongruent.' One trader described it to Bloomberg as 'a mini flash crash because of one headline.' Markets, as ever, price direction before destination.
"The firm does not dissolve. It reconfigures around a new axis of dependency—and the new landlord is the AI platform provider."
The Delegation Deficit: What Machines Still Need
There is a concept circulating among the consulting class: the delegation deficit. Most knowledge workers—including high-performing ones—cannot effectively hand cognitive work to an AI agent. They lack the ability to decompose goals into delegable sub-tasks, specify success criteria clearly enough for a model to self-evaluate, and maintain strategic coherence over a process they are no longer personally executing.
McKinsey documents the emerging arithmetic: two to three humans can now manage fifty to one hundred AI agents. A healthcare company replaced a ten-person software development team with three people—a product owner, an engineer who can effectively prompt AI tools, and a systems architect. BCG articulates the underlying logic: execution-level roles are automated first; the orchestration layer grows in complexity and importance, but demands a fundamentally different skill profile.
Compensation structures have not caught up. Payscale's 2026 report finds that 55% of companies offer no additional pay to employees who have demonstrably developed AI capabilities. The invisible premium on orchestration is real and growing—and most organisations have not yet figured out how to measure, reward, or recruit for it.
Three Brakes on the Inevitable
To describe these developments as inevitable is not to describe them as frictionless. Three constraints govern the pace.
Energy: a single AI agent query consumes approximately ten times the power of a conventional search. With inference costs now exceeding training costs at scale, the laws of thermodynamics are unimpressed by product roadmaps.
Opacity: when an autonomous agent makes a consequential decision and cannot explain its reasoning, the organisation deploying it inherits a liability it may not have priced. An ACM FAccT paper argues bluntly that black-box access is 'fundamentally insufficient for rigorous AI audits.'
Security: the OpenClaw framework's CVE-2026-25253 exposed over 17,500 internet-facing instances to unauthenticated access, leaking API keys for every major AI provider. As agents gain file-system and shell privileges, the attack surface scales with their usefulness.
The bottleneck in the agentic era is not compute. It is not energy, nor the black-box auditing problem, though both will occupy regulators for a decade. The bottleneck is the human being at the centre of the workflow who cannot quite bring themselves to let go of the wheel—not because the AI is untrustworthy, but because trust, once delegated, is very hard to claw back. That is not a technology problem. It is the oldest management problem in the world, now running at the speed of inference.
Sources: SWE-bench Verified (swebench.com); Payscale 2026 Compensation Best Practices Report; McKinsey & Company, 'The Agentic Organisation'; Gartner FutureScape 2026; CVE-2026-25253 / CVE-2026-27487 (NVD); Bloomberg; Barclays and Jefferies equity research notes; NBER, 'The Coasean Singularity'; ACM FAccT, 'Black-Box Access is Insufficient for Rigorous AI Audits.'
