Continuum
All essays
Essay

The economics of delegation

Why cheaper access to intelligence makes coordination quality more valuable.

October 2026

The best tool is often the one nobody has time to use. It may produce a better answer, but finding it, learning it and connecting it to the rest of the job costs more attention than the improvement is worth. People settle for familiar software because the alternative comes with work of its own.

We think agents can change that calculation. They can search for capabilities, compare results and use specialized tools without requiring a person to learn each one. As that work becomes more reliable, access to a better capability can require less additional human effort. Intelligence does not become free. The cost of putting it to work changes.

Our thesis is that this supports a simpler experience for the user and a more specialized market underneath it. The valuable businesses will either make delegation work or provide a capability that the delegating system cannot cheaply replace. Being another agent is not enough.

The hidden price of better software

Software has at least two prices: what the vendor charges and what the customer must do to get value from it. The second includes discovery, setup, training, switching between systems and checking the result. A product can be inexpensive and still be costly to use.

Consider a hypothetical procurement task. A specialist service might be best at checking supplier certifications, another at comparing contract terms, and a third at verifying delivery conditions. A person may use one familiar system rather than coordinate all three. The lost quality is the price of avoiding the coordination work.

An agent could absorb some of that work. It could determine which checks matter, call the services, reconcile disagreements and return a recommendation with the supporting evidence. The customer would still define the objective and approve consequential choices. But the customer's attention would no longer be required at every handoff.

This is the distinction behind the claim that the marginal effort of using state-of-the-art capability can approach zero. We mean additional human effort for a well-specified, reliably delegated task. We do not mean zero compute, zero latency, zero review or access to every best-in-class tool on every job. The agent still needs to know what better means and whether the improvement is worth its cost.

Agents do not tire, but they can waste work

A person will stop comparing tools because the next comparison is tedious. An agent does not face that particular limit. It can keep checking, retrying and comparing until a stopping rule tells it not to.

That persistence is useful only if it is directed. A system that keeps searching after it has enough evidence is spending money without improving the answer. A system that keeps retrying a flawed plan can produce more mistakes rather than a better result. The economic advantage is therefore not unlimited activity. It is choosing the next useful action and knowing when to stop.

Anthropic's guidance on building agents makes the tradeoff explicit: agentic systems can exchange additional cost and latency for better performance, and simpler workflows are often preferable. That is a technical observation with a business consequence. Customers should not pay for autonomy where a fixed procedure does the job more cheaply and predictably.

We would judge the product on cost per accepted result, including the human time needed to review and repair it. Cost per model call is an input to that measure. Calls completed, tokens generated and tasks attempted are not substitutes for it.

A simpler interface can support a more complex market

When people operate the tools themselves, each additional product asks for another share of their attention. That limits how finely software can specialize. A narrow capability may be excellent but too small to justify another login, another training session or another procurement process.

An agent-facing product could avoid part of that burden. It might offer a precise service through a machine-readable interface, with clear inputs, outputs, prices and limits. The customer sees the completed work; the agent deals with the component.

We expect this to create room for specialists whose economics were previously constrained by the cost of reaching and educating an end user. A difficult document check, a narrow scientific calculation or a verified source of industrial data could become useful inside many workflows without becoming a destination application.

But easier access also makes substitution easier. If an agent can discover a specialist, it can discover its competitor. If the output can be compared objectively, distribution through an agent may expose the product to continuous price and quality competition. Specialization is an opening, not protection.

The two places value can accumulate

One is coordination. A system that understands the objective, chooses appropriate capabilities and verifies the outcome can earn the customer's trust. Its advantage might lie in the quality tests it uses, its handling of exceptions or its ability to complete work across systems without creating a new burden for the user.

The other is a specialized capability that remains difficult to reproduce. That could be rights-cleared data, an expert-defined evaluation process, a deeply integrated operational service or a tool with demonstrably better results on a valuable class of tasks.

Neither position wins by default. A coordinator can be bundled into an existing platform. A specialist can be copied, commoditized or blocked by the platform through which it reaches customers. The question is what advantage remains when the same base model and broadly similar tool access are available to everyone.

We would be skeptical of a product whose entire advantage is selecting a model, arranging a few prompts and presenting the result attractively. Those choices can improve a product, but they are weak foundations for a durable business unless they lead to something harder to reproduce.

Trust is an economic variable

A delegated task does not save much time if the user must audit every step. Verification can become the new coordination cost.

METR's July 2025 randomized study offers a useful warning. Experienced open-source developers working in familiar repositories took longer when allowed to use the early-2025 AI tools studied, even though they believed those tools helped. The result was specific to the tools, developers and tasks in that study. It does not establish that AI slows all software work, or describe current agents. It does show why perceived productivity is an inadequate measure of actual value.

The same issue applies beyond coding. A polished answer can transfer work back to the customer through missing evidence, wrong assumptions or cleanup. We think reliable delegation requires explicit success criteria, visibility into what remains uncertain and boundaries on what the system may do without approval.

These requirements are sometimes treated as friction. They can be the mechanism that makes delegation economically worthwhile. A short, well-placed question is cheaper than an unauthorized action that someone must reverse.

The strongest objection: the platforms keep the savings

Large software platforms already own distribution, integrations and customer relationships. If they bundle competent coordination at a low incremental price, an independent coordinator may find that its technical advantage is too small to overcome the switching cost. Specialists may face a similar squeeze as platforms include acceptable substitutes.

Open interfaces could intensify that pressure rather than relieve it. More interchangeable components can produce a better market for buyers and a worse one for suppliers. Much of the value created by agents may become customer surplus rather than startup revenue.

We accept that risk. Our thesis does not require every layer to sustain attractive margins. It requires a subset of companies to own a measurable advantage that bundling does not erase: better outcomes on hard tasks, access that competitors cannot buy on the same terms, or operational depth that a generic platform cannot cheaply reproduce.

What we look for

We look for evidence that delegation removes work rather than hides it. Does the customer spend less time managing the task? Is the result accepted more often? What happens when inputs are incomplete, a service fails or two sources disagree? Does the system improve without asking users to become its supervisors?

For specialists, we ask why a rational agent would keep choosing the product after testing alternatives. For coordinators, we ask why a rational customer would keep paying after its existing platform offers a bundled version.

The signals that would weaken this thesis are equally concrete: review time fails to fall, exceptions keep requiring bespoke human intervention, or bundled products match standalone quality before the standalone companies develop a lasting advantage.

The change we expect is not a world in which people use infinitely many tools. It is a world in which they need to understand fewer of the tools doing their work. The durable business sits in the difference between assembling capabilities and delivering an outcome someone can trust.

Sources

These sources support the technical and empirical observations. Market structure and investment conclusions are Continuum's analysis, not forecasts established by the sources.