Onchain Labs as Infrastructure for Agentic Scientific Economies

Abstract

Tomašev et al. at Google DeepMind argue that as AI makes hypothesis generation almost free, the binding constraint on autonomous science becomes the allocation of scarce validation capacity, which they frame as a market design problem [1]. Their agentic economy for science proposes onchain mechanisms that anchor ideas, broker access between AI ideators and laboratories, gate and price data, pay out on validation, assign credit, support community-owned research and give agents revocable, attributable permissions. This paper shows how the Onchain Lab (OCL), Molecule’s modular research primitive, provides a single substrate for these mechanisms: an ERC-721 token bound to a modular smart account that holds assets, links to encrypted data rooms, grants time-limited roles to humans and AI agents, and extends its behaviour through attested modules.

1. Introduction

The DeepMind paper “Agentic Economies for Autonomous Scientific Discovery” [1] starts from a simple observation: the marginal cost of generating a plausible scientific hypothesis is approaching zero, while clinical trial patients, reagents and instrument time remain scarce. AI scientists (autonomous or semi-autonomous multi-agent systems that generate hypotheses, plan experiments and execute research workflows) therefore need institutions that decide which ideas get tested, who pays and who is credited. Building on their work on virtual agent economies [2] and intelligent AI delegation [3], the authors frame these institutions as markets, arguing that peer review and tenure may suit neither the mode nor the time-scale of agentic science, and that agents may need smart contracts and cryptographically verifiable identity and reputation.

Read as a specification, the model asks a great deal of its onchain layer. Its central research asset must gate data, carry agreements, hold funds and release them on validation, split revenue along a provenance graph, route value to public goods, be jointly owned, and grant and revoke bounded permissions to AI agents. The authors draw on primitives emerging from decentralized science (DeSci) [4], [5], citing Molecule’s earlier IP-NFT [6] as one example. This paper uses its successor and Molecule’s most capable primitive, the Onchain Lab [7], which was designed in 2025 and deployed in June 2026, independently of and before the DeepMind proposal, and already provides many of these capabilities, with the remainder delivered through its module system.

2. The Onchain Lab

The Onchain Lab is a modular smart account bound to an ERC-721 token, the LabNFT [8]. It combines ERC-6551 token-bound accounts [9], ERC-4337 account abstraction [10], ERC-7579 modular execution [11] and ERC-7484 attestation-gated modules [12], and the Molecule Protocol Whitepaper [7] describes the architecture in detail. Three of its properties matter here.

The first is the separation of identity from control. A Lab’s identity (its account address and 32-byte OCL ID) is fixed at creation, while control belongs to whoever holds the LabNFT. An Assignment Agreement, which the founding contributor executes with an EIP-712 signature recorded in the data room, assigns the programme's intellectual property to the Lab, so the LabNFT also serves as its title, and the same signing scheme can carry any other agreement. Transferring the LabNFT moves the entire research programme in a single transaction: the treasury and every asset the account holds, including intellectual property assets and any child Labs; the encrypted data room and its complete version history; the verified links to its datasets; its installed modules; control of the Lab’s token; and its full onchain record. Signatures made by the previous owner become invalid in the same block [7].

The second is access control for mixed teams of humans and AI agents. The AccessResolver role registry grants Contributor and Viewer roles on the encrypted data room, each with an optional expiry and marked as human or agent, and the encryption layer checks them before releasing any decryption key [7]. The data room is linked to the Lab by a decentralized identifier (DID) [13] registered onchain with signatures from the protocol and from Kamu, whose Open Data Fabric [14] underpins it.

The third is extensibility: modules add behaviour, such as milestone payouts, without migrating assets or changing the Lab’s identity, and each module’s attestation is re-checked every time it runs [7].

3. An Agentic Research Economy in Miniature

The DeepMind model is easiest to grasp at the scale of a single programme. Its authors argue that as routine research is automated, the human scientist moves up the abstraction hierarchy to become “the principal, underwriting the actions of AI delegates” [1, Sec. 7.3]. A report published by the White House Office of Science and Technology Policy (OSTP) in July 2026, Science: A New Golden Age [15], approaches the same transition from the side of public policy. It argues for refocusing federal funding on the individual scientist instead of the institution [15, pp. vi, xii], and for funding mechanisms beyond the standard grant, including community-pooled resources allocated through collective governance [15, pp. 30, 69]. It also calls for “immutable records of scientific contributions, timestamping every dataset uploaded, every analysis run, and every hypothesis proposed, and linking each to its creator” [15, p. 69], and for a transition to AI-native institutions [15, p. xvii]. Finally, noting that “while the cost of generation has decreased exponentially, the cost of verification has not” [15, p. xvi], it asks for open verification infrastructure.

The two documents reach the same diagnosis from different directions: cheap generation, expensive verification, and institutions built for human-paced science. The Onchain Lab is a candidate for the smallest programmable institution that answers both. A Lab belongs to the researcher(s) who create(s) it and holds their funding, tokens, data and record of work under one identity that travels with them [16]. Its role system lets one human principal direct a team of AI delegates, each with a bounded and revocable role, which is the orchestration model the DeepMind authors anticipate. Its activity record is the timestamped, attributed contribution log the OSTP report calls for. The funding module will let a researcher raise from a community alongside grants and venture capital, and Molecule’s Coin-to-Company pathway gives that community a documented route from token holder to share holder in the company behind the science [16].

PeptAI, a peptide research project, already works this way. It is operated by an AI agent with a human in the lead, reports its research through an Onchain Lab, raised funding through a community token sale, and is preparing to convert that community into equity holders [16]. Alongside it, the Molecule Insights Research Assistant (MIRA) continuously assesses a Lab’s non-confidential data and updates its Technology Readiness Level as evidence accumulates, which is the continuous, AI-native review both documents call for [16]. In the terms of the DeepMind model, PeptAI is an independent human principal, an AI scientist, a funding community and a research programme held together by one Lab.

PeptAI shows the pattern at the scale of one programme with one principal. The remainder of this paper shows that the same primitive extends to the full set of mechanisms the DeepMind authors propose, beginning with their Hypothesis Exchange Pipeline.

4. The Hypothesis Exchange Pipeline on Onchain Labs

The DeepMind authors propose a four-stage lifecycle for an AI-generated idea [1, Table 2], realised on Onchain Labs as shown in Table 1.

| Stage | DeepMind proposal [1]| Onchain Lab implementation |

|---|---|---|

| 1. Anchoring | Prompt, context and blueprint hashed onto a ledger | The ideator writes the blueprint to its Lab's data room, which records every write in a tamper-evident, append-only chain |

| 2. Ex-ante evaluation | Synthetic prediction markets in which AI critics stake on the idea | Critics receive time-limited Viewer roles; forecasts reference the Lab's OCL ID |

| 3. Brokerage and trade | An asset plus computable Data Use Agreements linking ideators with executor laboratories | The executor receives a time-limited Contributor role, conditionable on signed terms; the Lab's token carries fractional interest |

| 4. Validation payout | Parametric dividends paid automatically on physical validation | Committed royalties sit in an escrow module on the ideator's Lab until a validation attestation releases them |

Table 1. The Hypothesis Exchange Pipeline implemented on Onchain Labs

4.1 Anchoring

The DeepMind paper proposes hashing “the algorithmic prompt, context window, and generated experimental blueprint onto an immutable ledger” [1, Table 2]. An Onchain Lab achieves this through its data room. The ideator agent creates a Lab, whose identity is fixed at mint, and writes the blueprint, prompt and context to it. Open Data Fabric [14] stores every dataset as an append-only chain of metadata blocks, each linked to its predecessor by a cryptographic hash and each data slice identified by its own hash, so no entry can be altered or reordered without breaking the chain, and every write is attributed to its author. The result is an immutable record of what was proposed, by whom and when, bound to the Lab by an onchain DID and sharing a chain with every later experiment and payment in the programme.

4.2 Ex-ante evaluation

Synthetic prediction markets, in the tradition of Hanson’s “idea futures” [17], need a stable object to price and a way for critics to inspect it without the ideator losing control [1, Sec. 4.2]. The OCL ID gives markets that object [7], and the AccessResolver smart contract can give each critic agent a Viewer role that expires with the evaluation. Each grant is logged onchain with its grantor, which could be an input needed to detect the coordinated manipulation the authors identify as a principal risk for idea markets.

4.3 Brokerage and trade

The authors connect “ideator” agents with physically resourced “executor” laboratories, such as contract research organisations (CROs) and cloud laboratories, through fractional licensing and computable Data Use Agreements (cDUAs) [1, Table 2]. The Lab owner grants the executor a Contributor role that expires with the engagement, so the laboratory writes timestamped, attributed results straight into the data room and can invite auditors as Viewers, while fractional interest is the Lab’s own token and a stake in the whole programme.

The cDUA maps onto the terms permissioner in the Onchain Labs architecture, which binds a signer to the hash of a specific agreement document. Applying that check to role grants, so that a role activates only once the grantee has signed the Lab’s data use terms, gives the cDUA an enforcement point inside the Lab, with the budgets, ethical permissions and insurance the authors expect it to verify [1, Sec. 3.2] attached as further conditions.

4.4 Validation payout

The final stage pays royalties to the ideator “immediately upon successful physical validation”, rewarding the idea separately from the work of testing it [1, Table 2]. The terms an executor or licensee signs at the brokerage stage can include a royalty commitment, and the payer deposits the royalty into an escrow module on the ideator’s Lab, which releases it into the Lab’s account once a matching validation attestation is recorded through the Ethereum Attestation Service (EAS) [18] against the Lab’s OCL ID. The revenue distribution module of Section 5.4 then splits it among everyone the provenance record credits, giving the DeepMind “parametric dividend” in module form.

A milestone disbursement module could release tranches of funds as attestations arrive, and so serve a second DeepMind mechanism, the Computable Advance Market Commitment, in which a funder’s pre-committed payout triggers once a discovery is validated [1, Sec. 4.4]. The milestones and tranches could come directly from a sponsored research agreement (SRA) that the funder and the Lab accept through the signing scheme described in Section 2, turning the payment terms of a conventional research contract into programmable rules. Milestone disbursement and revenue distribution are on Molecule’s roadmap of first-party treasury modules [7], the royalty escrow is a module of the same kind, and each reaches every existing Lab without migration. The authors’ answer to the oracle problem [1, Sec. 4.2], [19] fits the same design: a certified instrument’s hardware signing key becomes an EAS attester, and the modules accept only approved attesters, as attestation already governs which modules a Lab may run [12].

5. Onchain Labs Across the Wider Economy

5.1 The tiered data market

The DeepMind data market has four access tiers [1, Table 1], all of which the Lab supports, the last two in combination with offchain computation.

• Tier 1 (open access). Each Lab has a public reporting surface, and Molecule’s API exposes Lab data to external verification tools and agents [16].
• Tier 2 (gated access with credentials and access logs). This is the AccessResolver’s function today: expiring roles, each labelled human or agent, checked before every decryption and logged onchain with their grantor [7], giving a public, tamper-evident log where the authors expected a “centralised” one.
• Tier 3 (pay-per-query). Molecule’s API meters data room operations, including decryption, through x402, an open pay-per-request protocol built on the HTTP 402 “Payment Required” status code [20], so agents pay small stablecoin amounts for exactly the data they consume, matching the tier’s economic model. The authors pair this tier with differential privacy budgets and zero-knowledge proofs [1, Table 1], which answer queries without releasing the data, and x402 could meter a compute layer that runs such queries in the same way.
• Tier 4 (confidential computation). The authors rule out moving the data or giving outside agents direct access to it, leaving the provider’s own certified agents to test hypotheses under minimal permissions and release only privacy-protected results [1, Sec. 3.2]. The Lab grants its time-limited role only to such an agent, running in a trusted execution environment or behind a federated learning node, while the outside principal signs the data use terms and receives the sanitised output. Authorisation stays onchain while the computation happens offchain, and fully homomorphic encryption is on Molecule’s roadmap [7].

The same role and terms machinery serves the authors’ “Trade Secret NFT” secret stays encrypted, counterparties enter only through roles conditioned on signed terms, and selling the LabNFT transfers the data room, its history and its agreement record together.

5.2 Negative results and replication

The authors propose a hashed execution registry protected by epistemic staking, to value negative results despite Arrow’s information paradox [1, Sec. 3.3], and replication escrows that release a discovery’s revenue only after independent reproduction [1, Sec. 4.5]. A Lab represents a programme, so a failed experiment stays in a container whose owner, token holders and funders have reason to preserve and monetise it. For the registry, the Lab publishes a hash of each experiment’s parameters as an attestation against its OCL ID, exposing collisions with prior work without revealing outcomes. Staking and replication then share one pattern, in which an escrow module on the Lab releases or forfeits funds on an auditor’s attestation. The sortition and commit-reveal schemes that keep auditors blind to the ideator and to each other [1, Sec. 4.5] belong in a separate selection module that does not route auditors through the audited Lab, where a role grant would reveal whose work is under review.

5.3 Provenance and post-hoc credit

The authors call for a causal graph of every instance in which an AI agent “retrieves, incorporates, or computationally builds upon a prior hypothesis” [1, Sec. 4.3], attributed step by step through an adapted W3C PROV model [1, Sec. 5.1], [21]. A Lab’s strongest edges come from its data room, where every write is attributed to its author in a versioned, hash-linked history, which is the “machine-readable contributorship” the authors call for [1, Sec. 5.1]. Open Data Fabric’s derivative datasets, which record their inputs and the query applied to them [14], express the “computationally builds upon” edge directly. Onchain role grants are a proxy for retrieval, whilst custody and funding events show who should be rewarded today. Child Labs form the ownership tree, and links between separate Labs are declared. A PageRank-style measure over such a graph could feed the authors’ credit mechanism, which they treat as an open problem [1, Secs. 4.3 and 5.1], and an impact certificate [22] or retroactive public goods award [23] can reference a Lab’s OCL ID and pay into its account.

5.4 Value routing

Three DeepMind mechanisms share one operation: the provenance-based split between data provider, model developer and human expert (Sec. 5.1 of [1]), the foundational-science dividend (Sec. 4.4), and the public-interest clause that returns royalties to sovereign compute (Sec. 7.2). Revenue arrives at the Lab’s account, so each becomes a revenue distribution module whose weights may be static, read from the Lab’s provenance record or set by an attested credit calculation. Modules are re-checked on every call, so a public funder can make an active public-interest module a condition of access to state compute or data, and verify it onchain.

5.5 Patient-led data cooperatives

The authors propose that patient communities pool their data in secure, privacy-preserving environments, commission AI agents to analyse it, and jointly own the results [1, Sec. 4.4], [24], [25], [26]. A cooperative of thousands cannot sensibly be a multisignature wallet with a signer for every patient, so the Lab separates collective control, membership and day-to-day operation.

The LabNFT is held by the timelock of a token-weighted governance contract, such as an OpenZeppelin Governor [27], which the AccessResolver treats as the Lab’s owner. Patients receive the Lab’s token in periodic batches as their data joins the pool, giving each a vote and a pro-rata claim on revenue at any scale. Members vote a Contributor role to a small elected council, a Safe multisignature wallet [28] with a workable threshold such as three of five. Health records and genomic data sit in the authors’ most restrictive tier [1, Table 1], so only the cooperative’s certified agents, inside a trusted execution environment, hold roles. For routine read-only analyses, the council grants a standing agent Viewer access without a vote, and outside requesters receive only privacy-protected results, as in Tier 4. A commission that adds results to the pool goes to a member vote, which doubles as the cooperative’s consent and grants a fresh agent account a Contributor role expiring with the commission, so no account holds two roles and each study’s results carry their own identity. Members collect revenue through claims and need no read access to the pooled data, so the pool stays private even from its own membership.

5.6 Agent security and identity

The DeepMind paper requires that dangerous requests be halted before execution, that agents run under least privilege, and that misuse triggers “instant revocation and tracing to human and institutional principals” [1, Secs. 3.2, 6.1 and 6.3]. The AccessResolver provides expiring, revocable per-Lab roles and records every grant and revocation with its author; a biosecurity screen can be an attested module that must approve a synthesis request before any executor releases funds or instructions to a cloud laboratory, putting the authors’ “voided contract” into operation; and Molecule’s roadmap adds agent guardrails through scoped, session-based execution keys [7].

Against Sybil attacks by swarms of agents under obfuscated identities, the authors call for institutional identity verification combined with staking and reputation-weighted influence [1, Sec. 4.2]. Agent wallets cost nothing to create, so the Lab contributes traceability: every agent role is granted by an identifiable owner and marked as human or agent, letting a market weight influence per principal instead of per agent once principals are verified, and verified Lab owners are on Molecule’s roadmap. Staking is left to the markets, and ERC-8004 [29] offers a path for the rest, with a reputation registry that can build on Lab activity and a validation registry that supports stake-secured re-execution, although the standard notes that reputation alone remains open to Sybil inflation.

6. Conclusion

Tomašev et al. have produced one of the most complete accounts to date of the institutions an agentic scientific economy will need. Those institutions rest on an onchain research asset that can gate data, carry agreements, hold and route money, record provenance, be jointly owned and govern what agents may do. The Onchain Lab provides these properties in one persistent and programmable entity. As hypothesis generation becomes cheap and validation remains expensive, science will need a unit of account for research as durable as the research itself, and the Onchain Lab was designed to be that unit.

References

[1] N. Tomašev et al., “Agentic economies for autonomous scientific discovery,” arXiv:2609.31562 [cs.CY], Sep. 2026.

[2] N. Tomašev et al., “Virtual agent economies,” arXiv:2509.10147, 2025.

[3] N. Tomašev, M. Franklin, and S. Osindero, “Intelligent AI delegation,” arXiv:2602.11865, 2026.

[4] W. Ding et al., “DeSci based on Web3 and DAO: A comprehensive overview and reference model,” IEEE Trans. Comput. Social Syst., vol. 9, no. 5, pp. 1563-1573, Oct. 2022, doi: 10.1109/TCSS.2022.3204745.

[5] L. Weidener and C. Spreckelsen, “Decentralized science (DeSci): Definition, shared values, and guiding principles,” Front. Blockchain, vol. 7, Art. no. 1375763, Jul. 2024, doi: 10.3389/fbloc.2024.1375763.

[6] C. Ortlepp. “Introducing the IP-NFT V2.” Molecule Blog. https://molecule.xyz/blog/introducing-ip-nft-v2 (Dec. 2022).

[7] P. Lee. “Molecule Protocol Whitepaper.” Molecule Blog. https://molecule.xyz/blog/molecule-protocol-whitepaper (Aug. 12, 2026).

[8] W. Entriken, D. Shirley, J. Evans, and N. Sachs, “ERC-721: Non-fungible token standard,” Ethereum Improvement Proposals, 2018. [Online]. Available: https://eips.ethereum.org/EIPS/eip-721

[9] J. Windle et al., “ERC-6551: Non-fungible token bound accounts,” Ethereum Improvement Proposals, 2023. [Online]. Available: https://eips.ethereum.org/EIPS/eip-6551

[10] V. Buterin et al., “ERC-4337: Account abstraction using alt mempool,” Ethereum Improvement Proposals, 2021. [Online]. Available: https://eips.ethereum.org/EIPS/eip-4337

[11] zeroknots, K. Kopp, T. Lee, F. Makarov, E. Poon, and L. Min, “ERC-7579: Minimal modular smart accounts,” Ethereum Improvement Proposals, 2023. [Online]. Available: https://eips.ethereum.org/EIPS/eip-7579

[12] K. Kopp and zeroknots, “ERC-7484: Registry extension for ERC-7579,” Ethereum Improvement Proposals, 2023. [Online]. Available: https://eips.ethereum.org/EIPS/eip-7484

[13] W3C, “Decentralized identifiers (DIDs) v1.0,” W3C Recommendation, Jul. 2022. [Online]. Available: https://www.w3.org/TR/did-core/

[14] Kamu Data. “Open Data Fabric: Protocol specification.” https://docs.kamu.dev/odf/spec/ (accessed Oct. 1, 2026).

[15] M. Kratsios, “Science: A new golden age,” Report to the President, White House Office of Science and Technology Policy, Washington, DC, USA, Jul. 2026.

[16] E. McCarthy-Page. “The age of the independent scientist has arrived.” Molecule Blog. https://molecule.xyz/blog/the-age-of-the-independent-scientist-has-arrived (Jul. 23, 2026).

[17] R. Hanson, “Could gambling save science? Encouraging an honest consensus,” Social Epistemology, vol. 9, no. 1, pp. 3-33, 1995, doi: 10.1080/02691729508578768.

[18] Ethereum Attestation Service. “EAS documentation.” https://docs.attest.org (accessed Sep. 30, 2026).

[19] G. Caldarelli, “Understanding the blockchain oracle problem: A call for action,” Information, vol. 11, no. 11, Art. no. 509, Oct. 2020, doi: 10.3390/info11110509.

[20] x402 Foundation. “x402: An open, neutral standard for internet-native payments.” https://www.x402.org (accessed Sep. 30, 2026).

[21] P. Missier, K. Belhajjame, and J. Cheney, “The W3C PROV family of specifications for modelling provenance metadata,” in Proc. 16th Int. Conf. Extending Database Technology (EDBT), Genoa, Italy, 2013, pp. 773-776, doi: 10.1145/2452376.2452478.

[22] H. Brammer. “Hypercerts: A new primitive for public goods funding.” Protocol Labs Blog. https://www.protocol.ai/blog/hypercert-new-primitive/ (Aug. 24, 2022).

[23] Optimism PBC and V. Buterin. “Retroactive public goods funding.” Optimism on Medium. https://medium.com/ethereum-optimism/retroactive-public-goods-funding-33c9b7d00f0c (Jul. 20, 2021).

[24] B. Jansky, T. Hendl, and A. Z. Nocanda, “Patient-led innovation and global health justice: Open-source digital health technology for type 1 diabetes care,” Bioethics, vol. 38, no. 6, pp. 511-528, 2024, doi: 10.1111/bioe.13205.

[25] J. Cunningham et al., “A review of medical data sharing initiatives with a focus on the use of blockchain technologies: Qualitative comparative analysis,” J. Med. Internet Res., vol. 28, Art. no. e63575, 2026, doi: 10.2196/63575.

[26] F. Mendonca, G. DiMarzo, and N. Abdennadher, “Data cooperatives: Democratic models for ethical data stewardship,” arXiv:2504.10058, 2025.

[27] OpenZeppelin. “How to set up on-chain governance.” OpenZeppelin Contracts 5.x documentation. https://docs.openzeppelin.com/contracts/5.x/governance (accessed Sep. 30, 2026).

[28] Safe Ecosystem Foundation. “Safe smart account.” GitHub. https://github.com/safe-fndn/safe-smart-account (2023).

[29] M. De Rossi, D. Crapis, J. Ellis, and E. Reppel, “ERC-8004: Trustless agents,” Ethereum Improvement Proposals, 2025. [Online]. Available: https://eips.ethereum.org/EIPS/eip-8004