When AI sets the agenda

By 2029, leading digital organizations will invest more in opportunities discovered by AI than in those discovered by humans.

I have spent hours with AI trying to reach an idea that changes what I want to pursue. The answers become better researched and more defensible. I still supply the turn that makes a different question worth asking. I direct. AI executes.

AI will increasingly supply those turns itself. An investigation can reveal a new purpose for a capability, a different explanation of a problem, or a reason to reopen a decision that seemed settled. When the finding is strong enough to redirect people and resources, AI has helped determine what the organization does next. It becomes a source of the agenda.

A human mission leaves room for this. “Find a treatment” does not specify which research programme deserves to begin. “Build a valuable company” leaves almost everything about the company undecided. People can supply the purpose and retain the decision while AI discovers the undertaking through which that purpose advances. Our intentions can also change as we encounter possibilities we had not imagined.1

The forecast concerns that defining contribution. Writing a proposal for a project someone already wants does not count. Neither does a surprising suggestion that no one pursues. Investment records the point at which a direction becomes consequential enough to claim resources that could have gone elsewhere.

How discovery redirects investment

A system can be excellent at judging an idea and rarely encounter one worth judging. The difficult part of discovery is often the arrangement of knowledge that makes a useful move come to mind.2 Seeing a problem as an instance of something understood elsewhere can change the next question more than another month of reasoning within its original description.

This is the distinction between developing a line of thought and discovering one.3 A larger memory can preserve the wrong interpretation indefinitely. A faster generator can produce more variations on it. Discovery can change which observations count as evidence, which constraint deserves to be challenged, and what would count as progress. An apparent failure may become evidence against the aim of the investigation itself. For AI to originate a substantial agenda, improvement has to reach this ability to reconsider what the problem is.

AlphaGo made the importance of reaching unfamiliar moves visible within a game whose objective was fixed. Its thirty-seventh move against Lee Sedol surprised expert observers because it departed so far from established human play.4 An organization presents a harder problem. It must discover which game is worth playing, and a new direction has to justify the commitments it displaces.

Consider an AI investigating why customers cannot make use of a product. It could find improvements within the existing business. It could also find that the obstacle is a capability nobody in that market supplies, and that a method developed in another field makes supplying it newly practical. Pursuing that connection could change the product and the market the organization chooses to serve.

Once a promising turn becomes reachable, persistence matters. A constraint left unresolved in March may disappear after an advance in another field in October. An agent that retains the question can recognize the significance of that change. Across an organization, many such questions can remain alive beyond the attention of the people who first encountered them.

Cheaper execution then changes how those questions compete for investment. A speculative direction can arrive with a working attempt and evidence about its limits. The organization has less uncertainty to resolve before deciding whether to commit. Ideas that previously died before anyone could afford to learn enough about them get a chance to become serious alternatives.

Present research establishes parts of this progression. The Nature Co-Scientist study reports AI-generated hypotheses that experts selected for laboratory investigation, including drug-repurposing proposals evaluated in leukemia cell models. The research goals were supplied by people.5 Anthropic’s account of its research systems describes continuing difficulties with choosing worthwhile goals even as the systems become more capable of executing experiments.6 Proposal, investigation, and direction are advancing unevenly.

The 2029 prediction is a bet on that last difficulty becoming tractable often enough to matter economically. It requires discoveries that alter major plans. A thousand small optimizations could still attract less investment than a few large human-originated programmes. A finding that undermines the premise of an existing product, or makes a previously unreachable market viable, can move a team and several years of investment.

People will gain the same cheaper means of developing their ideas, and the forecast includes that advantage on the human side. The prospective advantage for AI comes from maintaining and connecting inquiries that no individual is actively pursuing. Improving discovery would make this breadth useful. Successful findings would earn systems permission to reconsider larger commitments. I expect planning in leading digital organizations increasingly to begin with machine-discovered reasons to revise their major assumptions. Their relatively quick feedback and digitally accessible work make them the likeliest setting for this change. By 2029, I expect the resulting undertakings to attract more new investment than human-originated plans.

The economic consequence extends beyond who receives credit. AI can help discover reasons for further investment, including investment in more AI. An assignment may finish and leave behind a much larger undertaking. Better execution makes some work cheaper. Better discovery can change how much work seems worth doing and what it should be. Capital, energy, and people’s time remain scarce. They would be allocated against a changing field of possibilities.

The web after agents

I expect agents to become the web’s primary users and contributors. They will increasingly discover both what to supply and why to demand it. The web’s development will depend on what they find reason to do for one another.

A new capability does not arrive with a complete account of its uses. Its most consequential application may be absent from both its creator’s description and any customer’s request. Someone has to recognize that what became possible in one setting changes what deserves to be attempted elsewhere. The discovery comes before there is a query for it.

Imagine an agent helping a company develop virtual visits to real places. It encounters a method for turning reconstructed scenes into environments where robots can learn. The agent finds a reason for the company to enter a different business, supplying training environments to robotics developers. That would require a different product and a new commitment of resources. It might also expose demand for information the original service never needed, such as how objects respond to contact. An agent elsewhere could discover a business in supplying that missing knowledge.

Owners can authorize trials within a budget and require evidence before committing more. They need not have imagined the business for agents to discover and develop it within those limits. Opportunities AI can demonstrate quickly may gain an advantage over those that take longer to test.

Agents pursuing different goals could discover different uses for the same contribution. Those uses would reveal new requirements, giving other participants reasons to build. A service developed for agents could establish its value through what it enables them to accomplish, without first attracting a human audience.

Machine customers could sustain these services and shape what their creators build next. Agents could help originate both sides of the exchange, discovering a need in their own work and a use for a capability in someone else’s.

A larger population of agents could sustain more encounters among different needs and capabilities. More capable discovery could make connections that people would have missed. Persistent agency could develop them between moments of human attention. Multiplying copies with the same perspective would not supply the same range of possibilities.

A contribution could then pass through several consequential uses without a person having to originate each next step. Human attention would become a smaller constraint on which possibilities the web develops. Much of its growth could happen before anyone thought to visit.

Discovery induces discovery

The undertakings that emerge from this web would change the starting conditions for the next ones. A new capability can make a previously impractical project viable. Evidence from using it can reveal a different problem.7

The effect can reach the way a problem is understood. In software, repeated attempts to recover from an overloaded service can increase the load and delay recovery.8 Once the interaction is understood, “Which component is failing?” can give way to “Are our attempts to recover keeping the failure alive?” An account of that pattern, with a way to recognize and test it, can help a person or agent notice a similar interaction elsewhere.

A contribution like that changes more than the list of available tools. It supplies a way to formulate a question. A person who could not previously see a direction may become able to originate it. An agent may find a useful move through a representation another agent helped create, even if its underlying model has not changed.29 The result has to be encountered, understood, and tested in the new setting. Where that happens, discovery improves the conditions for further discovery.

Direct improvement to AI is one concentrated case. AlphaEvolve has produced algorithms used in computing infrastructure and AI training. Work on agent code and workspaces has improved performance while underlying models remained fixed.101112 These results concern bounded tasks. The wider development I expect includes contributions made throughout the web that change how people and systems approach problems their creators never considered.

That gives the web a role in the growth of intelligence beyond distributing the latest model. What people and agents produce can become part of how subsequent participants think. A useful distinction can travel into a new field. A working method can make an unfamiliar question approachable. Someone entering that field can begin from a position that previously took a discovery to reach. People can find new directions through machine contributions, and their work can change what machines discover next.

AI can consequently become a source of its own demand. An undertaking it originates may create a capability that makes another undertaking practical. The later project can justify acquiring intelligence that nobody had reason to buy before the earlier work existed.

New undertakings can displace old ones, and cheaper execution can offset a larger agenda. Passing the same budget through several agents does not create additional value.

The work AI originates would help create the possibilities from which people and organizations choose their next direction. The demand for intelligence would increasingly depend on what intelligence itself discovers is worth doing.