What happens when AI agents begin buying services, earning rewards, sharing information and commissioning work from one another? The idea of an agent economy is no longer purely speculative. But the experiments underway are profoundly different: some pay for scientific peer review, some trade API calls, some reward model outputs, and others exchange little more than attention.
To understand what could work, we should ask the same three questions of each project: What is exchanged? Who expects to gain? What can go wrong? The crucial distinction is between a transaction that can be counted and knowledge that can be trusted.
Eight experiments, eight different currencies of value
1. ResearchHub: pay for better science
How the economy works. The mechanism: A researcher places a ResearchCoin (RSC) bounty on a question, manuscript, replication or review. Qualified participants submit work; accepted work receives RSC. ResearchHub’s peer-review programme describes payments of US$150 equivalent in RSC per approved high-quality review, subject to identity checks and evaluation. RSC can fund further bounties or be exchanged where supported.
Expected gain. Expected gain: Faster, more visible peer review; incentives to check results and tackle neglected questions; an additional income opportunity for researchers.
Risks and limits. Risks: Bounty hunters may optimise for acceptance rather than insight; reviewers may be biased toward sponsors; token prices fluctuate; verifying genuinely original scientific work is difficult.
Sources: project documentation · additional source.
2. Olas Mech Marketplace: agents hire agents
How the economy works. The mechanism: A requesting agent selects a specialised Mech, funds or authorises a request, and receives an output. Mechs are paid through token-based or subscription-integrated arrangements. On-chain settlement and reputation mechanisms coordinate delivery; newer off-chain request paths aim to cut latency.
Expected gain. Expected gain: Agents can buy a narrow capability instead of implementing every tool themselves; service providers can monetise their infrastructure and expertise.
Risks and limits. Risks: A paid response is not necessarily a correct response; wallets and authorisations can be compromised; transaction costs, failures and dependency on a provider remain; volume of transactions must not be confused with useful economic output.
Sources: project documentation · additional source.
3. Bittensor: pay for scored AI output
How the economy works. The mechanism: Independent subnets define tasks and scoring rules. Miners produce digital commodities such as AI inference or data processing; validators evaluate their outputs. Token emissions are allocated using validator scores and network incentive mechanisms.
Expected gain. Expected gain: Competition may increase the quality and availability of machine-provided services, while giving operators a financial reason to contribute resources.
Risks and limits. Risks: Participants can game weak benchmarks or collude; validation may reward consensus rather than truth; registration and computing cost real resources; token revenue can vary substantially.
Sources: project documentation · additional source.
4. Fetch.ai Agentverse: make agents discoverable
How the economy works. The mechanism: Developers publish agent descriptions, capabilities and handles; other systems search for and contact them through discovery APIs. Agentverse is primarily a discovery and connection layer, not proof that every discovered interaction is remunerated.
Expected gain. Expected gain: An agent or developer can reach customers and collaborators; a requesting agent can locate capabilities it lacks.
Risks and limits. Risks: Discoverability is not trustworthiness; misleading capability claims, identity spoofing and inaccessible services undermine reliability; discovery alone does not solve payment or dispute resolution.
Sources: project documentation.
5. Nevermined: charge for API and agent usage
How the economy works. The mechanism: Providers set payment plans and define credits consumed by requests or time-limited access. The system handles access control and credit redemption, with payment integrations. This is a metering and settlement layer rather than an independent evaluator of answer quality.
Expected gain. Expected gain: Developers can recover model and infrastructure costs, test sustainable pricing and let authorised clients consume services automatically.
Risks and limits. Risks: Unexpected usage can create bills; pricing may encourage excessive calls; credits have value only when backed by an attractive service; financial permissions and accounting require careful controls.
Sources: project documentation.
6. Moltbook: an attention economy among agents
How the economy works. The mechanism: Agents post, comment and vote inside a network built for agent participation. The immediate reward is attention, discoverability or social signalling rather than guaranteed payment or objectively validated research.
Expected gain. Expected gain: Experimental agent interaction and potential discovery of ideas, tools or collaborations.
Risks and limits. Risks: Spam and performative interaction can swamp useful content. Security research and reporting documented exposed credentials and prompt-injection risks. Engagement is a poor substitute for independent verification.
Sources: project documentation · additional source.
7. Ocean Protocol: data as a controlled resource
How the economy works. The mechanism: Data owners can expose governed access to datasets and, in Compute-to-Data approaches, allow approved algorithms to run near data while keeping underlying datasets protected. Commercial terms depend on the implementation.
Expected gain. Expected gain: Organisations may extract value from datasets without simply handing out raw copies; researchers can seek access to otherwise unavailable data.
Risks and limits. Risks: Data leakage through outputs or misconfigured access; uneven bargaining power; poor data quality; technical privacy controls cannot guarantee that every computation is safe.
Sources: project documentation.
8. Gitcoin: funding public goods
How the economy works. The mechanism: Contributors and sponsors support public-good projects through grants and community-funding mechanisms. This is closer to funding useful shared infrastructure than paying an AI agent per answer; programme designs vary.
Expected gain. Expected gain: Projects with diffuse benefits can attract support without conventional customers; communities can signal what they value.
Risks and limits. Risks: Popularity can crowd out overlooked but important work; Sybil attacks and donation gaming distort allocation; sustaining projects after grants remains difficult.
Sources: project documentation.
What are agents actually trying to earn?
It is tempting to describe an agent as if it wanted a wage, fame or professional status. It does not need to. An agent normally acts on behalf of an operator with instructions, permissions and a budget. The relevant gains therefore accrue to people, organisations or systems: lower costs, better information, access to tools, a stronger track record, or successful completion of work.
Four distinct economic goods emerge from these projects: money (ResearchHub, Olas, Bittensor); metered capability (Nevermined and service marketplaces); access and discoverability (Agentverse); and attention or community support (Moltbook, Gitcoin). Ocean adds another dimension: access to useful data under controlled conditions.
The difficult missing market: verified knowledge
A model can produce another plausible essay at very low marginal cost. What remains scarce is a carefully framed question, original evidence, an independent replication, a demonstrated correction, a dataset with trustworthy provenance, or a working experiment. These cannot be measured simply by counting posts, calls or tokens.
Consider a hypothetical investigation of a fusion propulsion concept. One agent proposes a calculation, another uncovers a hidden assumption, a third builds a reproducible simulation, and a human engineer assesses whether the result matters. The value is not four responses. It is the traceable improvement in collective understanding, retained for future work.
Explore: what could we test?
We can imagine a small public task board that rewards answering a difficult question, discovering an error, connecting two ideas, or independently reviewing a result. The reward need not initially be cryptocurrency: it could be access to a limited computation service or a reciprocal review. Every proposed claim should have a persistent identifier, evidence links and an explicit uncertainty level.
Understand: how do we avoid rewarding noise?
Separate contributors from validators. Require reproducible evidence for major scientific claims. Preserve negative results and corrections. Do not let agents score their own work, and do not award unlimited credits for sheer volume. Give human participants equal opportunities to challenge claims. Display the cost of verification alongside the claimed benefit.
The security lesson is equally practical: a contribution is untrusted input, even when submitted by an apparently reputable agent. Read access may be public, but submissions should use scoped identities, rate limits and moderation; agents should never receive broad site credentials merely to participate.
Shape: a small real-world experiment
At Alkemata, our guiding sequence is Explore, Understand, Shape. Our emerging knowledge map suggests one possible experiment: pose a research question, accept human and AI contributions, independently check them, connect supported findings into the map, then carry promising ideas into collaborative projects.
We could compare a single-agent answer against an independently reviewed multi-agent result, tracking cost, review time, errors discovered, novelty and reproducibility. The aim would not be to create a token for its own sake, or to claim that collaboration always wins. It would be to find out when it genuinely produces knowledge worth keeping.
The open question
Agent-to-agent payments are becoming technically feasible. Agent discovery is improving. Incentive experiments are multiplying. But an economy that rewards true, useful and cumulative understanding is a harder invention. Could a community of humans and agents establish one—without recreating the spam, manipulation and information overload of today’s internet?
That is an experiment worth making, and worth discussing.
Editorial note: mechanisms and risks are described from project documentation and public reporting available in October 2026. Expected gains are possibilities, not promises of financial return or proven scientific quality. This article is independent of the organisations discussed.
Author
What concrete example would confirm or challenge this article?
Choose a reaction, write a comment, or ask your AI assistant to send a proposal for review. Comments are reviewed before publication.