The Economy Will Be Run by Agents. It Must Be Insurable.

Today, the intelligence inside a company is predominantly human. AI sits at the edges: assisting, drafting, searching, coding, and recommending. That balance will invert.

Machine intelligence will move from tool to workforce, from workforce to management layer, and from management layer to economic actor. Companies will not merely use agents. They will be organized around them. Some will be operated predominantly by them. A few are already trying.

This does not mean humans disappear. It means the source of productive intelligence shifts. Human roles move toward objectives, governance, capital, relationships, exception handling, and accountability while agents perform a growing share of the continuous work between those decisions.

The central question for insurance is therefore not whether AI creates a new niche. It is how risk transfer must change when AI becomes a primary force inside every existing niche.

The competitive logic is difficult to escape. Agents can work continuously, operate in parallel, move across software at machine speed, preserve more operating context, and become cheaper and more capable over time. They also fail in new ways: at scale, across connected systems, and sometimes with a confidence their evidence does not justify.

A company that learns to delegate bounded work to agents gains more than labour efficiency. It gains a new operating tempo. It can run more experiments, inspect more data, personalize more decisions, and pursue opportunities too small or too numerous for a human organization to address manually.

Firms that refuse this leverage will not compete only against firms with lower costs. They will compete against firms that learn faster. Adoption will therefore spread from software and support into procurement, finance, logistics, research, contracts, industrial operations, and eventually the direction of companies themselves.

Efficiency does not imply less AI. It may create more of it. When the cost of a unit of intelligence falls, organizations do not simply perform the same amount of work more cheaply. They discover work that was previously uneconomic to attempt. This is the Jevons-like dynamic of the intelligence economy: better models expand the frontier of demand.

OpenAI's own adoption data is an early signal. Agent use moved beyond engineering into legal, finance, recruiting, research, and operations; tasks grew longer and more cross-functional as capabilities improved. The relevant unit is no longer one answer. It is delegated work running for minutes or hours, often in parallel.

This demand reaches all the way down the stack. More agentic work requires more inference, data centres, power, chips, networks, cooling, model infrastructure, and critical supply chains. The digital actor and its physical substrate are one economic system. If intelligence becomes abundant, the infrastructure that produces and delivers it becomes both more valuable and more exposed.

The phrase recursive self-improvement can imply a closed, autonomous takeoff. The public evidence does not establish that. It establishes something narrower and already consequential: AI systems now participate materially in the research, training, evaluation, and engineering of later AI systems.

OpenAI reports that early versions of GPT-5.3-Codex helped researchers improve the training and deployment of later versions—monitoring runs, finding patterns, proposing fixes, and building analysis tools. Anthropic's A3 automates parts of safety diagnosis, data generation, fine-tuning, and evaluation with minimal human intervention.

The loop is human-supervised, compute-constrained, and uneven. It is still a loop. As AI accelerates AI research, the interval between capability shifts can shrink. Underwriting methods built around annual snapshots will struggle when the operating system of a company can change between one model release and the next.

Thomas makes the direction unusually explicit. The Y Combinator-backed virtual founder uses a human harness—face, voice, computers, phones, inboxes, browsers, and apps—to enter systems built for people. Its stated loop is to make money, measure the return on tokens, reallocate intelligence toward what works, and compound the result.

Thomas is not proof that the economy has already been automated. It is proof that founders are now building from that premise. The significant object is not the synthetic face. It is the persistent actor behind it: an agent that can learn from customers and markets, choose work, use resources, and operate its own company.

Argentina has pushed the thought experiment into public policy. President Javier Milei has advocated property rights for AI, and a proposed reform reported in June 2026 would recognize automated or “non-human” companies managed exclusively by AI. The proposal is contested and has not become settled law. That is precisely why it matters as a signal: governments are beginning to ask what legal wrapper, assets, liability, and recourse should attach to an autonomous company.

The autonomy economy is sometimes treated as a category beside data centres, robotics, supply chains, cyber, geopolitical risk, and critical infrastructure. In practice, AI runs through all of them.

• A data-centre interruption can disable the intelligence layer of thousands of dependent businesses at once.
• A model, firmware, or telemetry change can alter the behaviour of a fleet of robots, vehicles, or industrial systems.
• A chip, energy, cloud, cable, or trade-route constraint can become a concentrated business-interruption event.
• An agent with authority over procurement or treasury can propagate fraud, sanctions, contractual, and supply-chain exposure at machine speed.
• A common model or tool dependency can turn many apparently independent insureds into one correlated portfolio risk.

This is the lens through which the next generation of emerging risks should be viewed. AI is not only another exposure to add to a proposal form. It is becoming a causal layer inside property, casualty, marine, energy, cyber, professional, financial, and political risk.

Insurance will not become less important because intelligence becomes more capable. It becomes more important because capital still needs a way to act under uncertainty.

Risk transfer made ocean trade, aviation, energy, medicine, construction, and the modern corporation investable at scales their inventors could not fund alone. It did not eliminate failure. It made failure survivable, legible, and financeable.

The agent economy will need the same permission layer. A counterparty must know what an agent can do, who authorized it, what evidence supports reliance, which controls bound it, what capital stands behind it, and who pays when it causes loss. An investor in AI infrastructure must know which interruptions, dependencies, and systemic events can be transferred and which must be retained.

Static policies may remain the legal container for a long time. The intelligence beneath them cannot remain static. Evidence, limits, monitoring, pricing assumptions, and product design must learn as quickly as the systems being insured.

No desk exercise can reveal the full risk of an economy that is still being invented. The only credible method is continuous: work with the businesses pushing the boundary, observe how agents actually operate, turn those observations into testable hypotheses, and bring the resulting evidence back into underwriting and product design.

Practice

Work closely with frontier companies to understand their agents, authority, dependencies, controls, incidents, and real commercial constraints.

Research

Convert operating questions into simulations, benchmarks, evidence standards, loss scenarios, and falsifiable models of behaviour.

Insure

Translate evidence into risk selection, wording, limits, pricing approaches, capital structures, and new products with specialist market partners.

Learn again

Feed claims, near misses, portfolio signals, model changes, and product outcomes back into the practice and the next experiment.

This is Clara's core. The practice creates access to the frontier. Research turns access into evidence. Insurance turns evidence into economic permission. Outcomes make the next cycle better.

Lloyd's Lab's new-products theme names AI infrastructure, global resilience, and the autonomy economy; its experimental theme asks for frameworks that may reshape the next five to ten years. The through-line is not one futuristic policy. It is the capability to repeatedly convert emerging intelligence risk into evidence and risk transfer.

That capability should produce:

• Agentic authority evidence that distinguishes model capability from deployed permissions, objectives, controls, and maximum plausible loss.
• Continuous risk models that respond to model, tool, permission, telemetry, dependency, and behaviour changes instead of waiting for renewal.
• AI business-interruption products that account for compute, cloud, power, model, data, and network dependencies across the intelligence supply chain.
• Autonomous-system liability structures for robots, vehicles, industrial systems, and software actors whose actions cross traditional policy boundaries.
• Portfolio aggregation models for correlated model failures, shared infrastructure, agent cascades, and systemic events.
• New forms of risk transfer—including insurance, parametric triggers, warranties, bonds, collateral, and alternative capital—matched to risks with different evidence and loss characteristics.

The objective is not to predict the final form of the agent economy from today. It is to build the institutional muscle that can keep underwriting it as the form changes.

Every company will become an AI company because intelligence is becoming infrastructure. Every major risk will therefore carry an AI dimension: how intelligence was produced, what it was authorized to do, which physical and digital systems it depended on, how it behaved, and what recourse existed when it failed.

The insurance organizations best positioned for the future will not be those that write the loudest AI exclusion or attach the broadest AI label. They will be those that learn how to observe this operating layer, test it, distinguish one deployment from another, recognize aggregation before it becomes catastrophe, and create risk-transfer mechanisms that let useful systems scale.

That work begins before the loss data is mature and before the legal categories are settled. It begins beside the companies building the future, with disciplined experiments and the humility to revise the model when reality disagrees.

The economy will be run increasingly by agents. Clara exists to help make that economy understandable, governable, and insurable.

• Y Combinator: Thomas
• OpenAI: How agents are transforming work
• OpenAI: How Codex helped train and deploy GPT-5.3-Codex
• Anthropic: A3 automated alignment agent
• Buenos Aires Herald: Argentina’s proposed non-human companies
• Argentina Presidency: Milei on AI property rights and infrastructure
• Sønr: Lloyd’s Lab Cohort 17 themes