Inverting The Human-Agent Relationship

In 1997, IBM’s Deep Blue defeated then-World Champion Garry Kasparov in a chess match, and it became clear that computer chess engines would soon surpass humans. Interestingly, a well-prepared human working together with a computer—an arrangement often called a "centaur"—could outperform the strongest engines of that era.

Skilled human intuition could guide the engine’s search, navigate complex middlegames, and recognize nuances that standard engines missed. Combined with a computer’s brute-force calculation, the pair often led to better practical decisions than a computer alone.

As I think about AI systems impacting labor markets and economies over the next few years, I expect to see a similar pattern emerge. Agent systems will unleash countless units of intelligence on unsolved problems in the world, but they will not be able to do this without strong guidance and support from humans. Humans will guide the search space and help ask the right questions, allowing the AIs to work towards the answers.

The working assumption today is that agents will act on behalf of humans. While this is practical, and inevitable, the far more interesting economic unlock emerges when humans work for agents. In the next 24 months, I expect to see the first Zero-Employee Company, a concept my partner Kyle coined in his section of Frontier Ideas for 2025. Specifically, I expect the following:

1. A token-governed agent will raise over $1B to solve an unsolved problem (like curing a rare disease, or manufacturing nanofibers for applications in defense).
2. That agent will distribute over $100M in payments to humans (who are working for the agent to achieve the agent’s objectives in meatspace).
3. A new, dual-class token structure emerges that separates ownership based on capital and labor (such that financial incentives are not the sole input into overall governance).

Because agents are not anywhere near being both sovereign and able to handle long term planning and execution, in the short term, agents need humans more than humans need agents. This will yield novel labor markets that enable economic coordination between agent systems and humans.

The famous Marc Andreessen quote, “The spread of computers and the internet will put jobs in two categories: people who tell computers what to do, and people who are told by computers what to do,” rings more true today than ever before. I expect that in the rapidly evolving agent/human hierarchy, humans will serve two distinct roles — labor contributors who perform small, bounty-style tasks on behalf of agents, and a decentralized board of directors to provide strategic input in service of the agent’s north star.

This essay explores how agents and humans will co-create, and how crypto rails will provide the ideal substrate for this coordination, by examining three guiding questions:

1. What are agents useful for? How should we classify agents based on the scope of their objectives, and how does the range of human input required vary across those classifications?
2. How will humans interact with agents? How does human input — tactical guidance, contextual judgment, or ideological alignment — incorporate into the workflows for these agents (and vice versa)?
3. What happens as human input diminishes over time? As agents’capabilities improve, they become self-sufficient i.e. capable of reasoning and acting on their own. In this paradigm, what role will humans play?

The relationship between generative reasoning systems, and the people who benefit from them, will evolve drastically over time. I examine this relationship by looking forward from where agent capabilities are today, and working backwards from the endstate of a Zero-Employee Company.

What are agents useful for today?

The first generation of generative AI systems—the chatbot-based LLMs of the 2022-2024 era, like ChatGPT, Gemini, Claude, Perplexity, etc.—are largely tools designed to augment human workflows. Users interact with these systems via input/output prompt pairs, parse the responses, and then ascribe their own judgment as to how the results should be brought into the world.

The next generation of generative AI systems, or “agents,” represent a new modality. Agents like Claude 3.5.1 with “computer use” and OpenAI’s Operator (ie, agents that can use your computer) have the ability to directly interact with the internet on behalf of users, and can make decisions on their own. The key difference here is that the judgment — and ultimately the action — is being exercised by the AI system, not the human. The AI is taking on a responsibility previously reserved for humans.

This transition introduces a challenge: lack of determinism. Unlike traditional software systems or industrial automation, which function predictably within defined parameters, agents rely on probabilistic reasoning. This makes their behavior less consistent across identical scenarios and introduces elements of uncertainty — not ideal for critical situations.

Said another way, the existence of deterministic- vs. non-deterministic agents naturally bifurcate two ways to classify agents: ones that are best at scaling existing GDP, and ones that are better suited for creating new GDP.

1. For agents that are best at scaling existing GDP, the jobs are, by definition, already known. Automating customer support, handling freight forwarding compliance, or reviewing github PRs are examples of well-defined, bounded problems where agents can map responses to a set of expected outcomes in a straightforward way. The lack of determinism is generally bad in these domains, as there is a known answer; creativity is not needed.
2. For agents that are best at creating new GDP, the job is to navigate high uncertainty and unknown problem sets to fulfill long term objectives. The results here are much less straightforward, because there is inherently no set of expected outcomes for the agent to map to. Examples here include drug discovery for rare illnesses, breakthroughs in material science, or running net new physics experiments to better understand the nature of the universe. The lack of determinism can be helpful in these domains, as lack of determinism is a form of generative creativity.

Agents focused on existing GDP applications are already unlocking value. Teams like Tasker, Lindy, and Anon are all building infrastructure aimed at this opportunity. However, over time, as capabilities mature and governance models evolve, teams will shift their focus on building agents capable of solving the problems at the frontier of human knowledge and economic opportunity.

This next cohort of agents require exponentially more resources precisely because their outcomes are uncertain and unbounded — these are what I expect to be the most compelling Zero-Employee Companies.

How will humans interact with agents?

Agents today lack the ability to perform tasks that require physical interactions with the real world (driving a bull dozer) and/or that require human-in-the-loop tasks (sending a bank wire).

For example, an agent tasked with the mission of identifying and extracting lithium may excel at processing seismic data, satellite imagery, and geological records to identify promising sites, but fail when attempting to source the data and imagery itself, or resolve ambiguities in interpretations, or obtain permits and contract labor to run the actual extraction process.

These limitations require humans to augment agents as enablers to provide the necessary real-world touch points, tactical interventions, and strategic input required to complete said tasks. As the relationship between humans and agents evolves we can also distinguish the roles that humans play in agent systems:

1. First, labor contributors who operate on behalf of agents in the real world. These contributors help agents move physical things, represent the agent when a human is needed, perform work that requires arms and legs, grant access to experimental labs, or logistics networks, etc.
2. Second, a board of directors that provide strategic input and refine the local objective functions that drive agents’ day-to-day decisions while ensuring they align with the north star that defines the agent’s purpose.

In addition to these two, I also expect humans to play the role of capital contributors, funding agent systems with resources to go about and achieve their objectives. This capital will naturally come from humans at first, and other agents over time.

As agents mature, and the number of labor and guidance contributors grows, crypto rails provide the ideal substrate for human-agent coordination—especially in a world where agents are directing humans who speak different languages, get paid in different currencies, and live in different jurisdictions all around the world. Agents will ruthlessly pursue cost efficiencies and exploit labor markets in an effort to achieve their stated mandate. Crypto rails are necessary to equip them with a way to coordinate those labor and guidance contributors.

Recent crypto-enabled AI agents like Freysa, Zerebro, and ai16z represent simple experiments in capital formation — something we have written extensively about as a core unlock of crypto primitives and capital markets across a variety of contexts. These are the toys that will pave the way for an emergent new model for resource coordination, which I expect to occur as follows:

1. Step 1: Humans collectively raise capital via tokens (Initial Agent Offering?), establish a broad objective function and guardrails to inform the intended purpose of the agent system, and then assign control of the raised capital to the system (for example, new molecule development for precision oncology);
2. Step 2: The agent reasons about steps to allocate that capital (how to narrow the search space for protein folding and how to budget for inference workloads, manufacturing, clinical trials, etc.), and defines actions for human labor contributors to complete on its behalf via bespoke bounties (for example, input the set of all relevant molecules, sign a compute SLA with AWS, and run wet-lab experiments);
3. Step 3: As the agent runs into roadblocks or forks, it solicits strategic input from the board of directors when necessary (incorporate new papers, shift research method), allowing them to guide the behavior of the agent at the edges; and
4. Step 4: Eventually, the agent improves to a point where it is able to define actions for humans with increasing precision and requires minimal input on how to allocate resources. At this point, humans are only needed to align the system ideologically and prevent it from acting out of line with its initial objective function.

In this example, crypto primitives and capital markets give agents three critical pieces of infrastructure to attain resources and expand capabilities: first, global payment rails; second, permissionless labor markets to incentivize labor and guidance contributors; third, asset issuance and trading infrastructure, which is necessary for capital formation and downstream ownership and governance.

What happens as human input diminishes?

In the early 2000s, chess engines improved dramatically. Through advanced heuristics, neural networks, and increased compute, they became practically flawless. Modern engines like Stockfish, Lc0, and AlphaZero variants are so far beyond human capability that human input rarely adds value, and in most cases introduces errors that the engine would not make on its own.

A similar trajectory could unfold with agentic systems. As we refine these agents through iterative back-and-forth with human collaborators, it is conceivable that in the long run, agents become so competent and aligned with their objectives that the value of any strategic human input trends toward zero.

In such a world, where agents consistently navigate complex problems without requiring human intervention, the role of humans risks diminishing to that of passive observers. This is the core fear of AI doomers (however, it is not clear that this outcome is possible).

We are standing on the precipice of superintelligence, and the optimists among us would prefer that agent systems remain extensions of human intent rather than entities that evolve their own objectives or run with their autonomy unchecked. In practice, this means that human identity (personhood) and judgment (power and influence) must remain central to these systems. Humans will need strong ownership and governance rights over these systems, in order to ensure they retain oversight and anchor these systems in collective human values.

Picks and shovels for our agentic future

Breakthroughs in technology lead to nonlinear rates of economic progress, and the systems around them often break before the world adjusts. Agent systems are improving in capabilities at rapid pace, and crypto primitives and capital markets are already serving as the much needed coordination substrate for both advancing the buildout of these systems and setting guardrails as they embed themselves into society.

In order for humans to provide tactical support and active guidance for agent systems, we expect the following picks-and-shovels opportunities to emerge:

1. Proof-of-agenthood + proof-of-personhood: Agents lack the concepts of identity or property rights. As proxies for humans, they are reliant on human legal and social structures for agency. To bridge this gap, we need robust identity systems for both agents and humans. A digital certification registry could enable agents to build reputation, accumulate credentials, and interact with both humans and other agents transparently. Similarly, proof of personhood primitives like Humancode and Humanity Protocol provide strong guarantees of human identity against adversarial actors in these systems.
2. Labor markets and off-chain verification primitives: Agents need to know that tasks they assign are completed as per their objectives. Tools that allow agent systems to create bounties, verify their completion, and distribute payments are table stakes for any meaningful economic activity mediated by agents.
3. Capital formation and governance systems: Agents need capital to solve problems, and checks and balances to ensure they are acting in line with their defined objective functions. Novel structures to source capital for agent systems, and new forms of ownership and control that blend financial stake and labor contribution are going to be a rich design space in coming months.

We are actively looking to invest in these critical layers of the human-agent coordination stack. If you are building in this space, please reach out to us.