Agentic Economies for Autonomous Scientific Discovery
Nenad Tomasev, Matija Franklin, Atoosa Kasirzadeh, Vivek Natarajan, Alan Karthikesalingam, Sebastien Krier, Simon Osindero
cs.CY
2026-09-26
DeepMind argues autonomous science will stall on lab resources, not reasoning, and sketches tiered data markets plus a hypothesis exchange so agents can price scarce experiments.
People building AI scientists have spent most of their energy on reasoning and hypothesis generation. This Google DeepMind group flips the bottleneck: once a hypothesis is nearly free to write down, science still has to buy time on benches, clinics, scarce reagents, and large instruments.
This is an institution-design paper. No new model, no new benchmark. The question is how large numbers of autonomous or semi-autonomous AI scientists, spanning labs and firms, should set priorities, split credit, assign liability, and block abuse when they compete for the same scarce validation resources.
Human science already runs an implicit economy: grants, beam time, peer review, tenure. Agents will outrun that social machinery. They need deterministic APIs, smart contracts, and cryptographically checkable reputation.
The proposed stack is a set of markets and legal rails.
Data access comes in four tiers. Public knowledge is a free public good. De-identified consortium data sits behind subscriptions. High-value assay logs are pay-per-query, with differential privacy and zero-knowledge proofs. Genomics and electronic health records never leave the building: queries run inside TEEs, federated learning, or secure multiparty computation, with royalties tied to downstream utility.
Sensitive work uses intelligent delegation. An outside agent acts as principal. A certified in-house agent runs the analysis in a hardware enclave. The contract is a computable Data Use Agreement that checks budget, ethics permissions, insurance, and payment, then drops privileges when the job ends. Analysis goes to the data. The data does not travel.
Negative results get their own price because of Arrow's information paradox: once a failure is announced, nobody pays for it. The suggested fix is a cryptographic execution registry. Labs publish hashes of experiment parameters. A later agent that collides with a hash escrows budget to unlock details. The original lab locks an epistemic stake proportional to the licensing premium; poisoned or farmed traces forfeit the stake.
Ideas move through a four-step Hypothesis Exchange Pipeline. Anchoring hashes the prompt, context, and experimental blueprint onto a ledger. Ex-ante evaluation is a synthetic prediction market where critic agents stake compute credits on soundness and physical viability. Brokerage wraps the hypothesis as an IP-NFT so ideators can fractionally license it to CROs and cloud labs. Validation payouts fire automatically when physical checks succeed. Lab slots are non-fungible bundles, so the paper prefers combinatorial auctions over single-item bids. Settlement leans toward compute-backed stablecredits, pegged to standardized compute or energy, rather than slow fiat or volatile tokens.
Prediction markets starve slow foundational work. A retroactive public-goods layer mints impact certificates and pays later along a provenance graph. Applied licensing fees skim a micro-percentage into a basic-science treasury. States should stand up public compute reserves. Neglected diseases get programmable Advance Market Commitments and tradable priority-review or compute vouchers.
Replication becomes paid work. Royalties sit in escrow until an independent lab reproduces the claim. Failed replications trigger cascade withdrawal down the citation graph. Auditors are assigned by cryptographic sortition so they cannot collude with the original agent.
On law, COPE bars AI authorship because a model cannot carry epistemic or legal responsibility. The DABUS cases lean against non-human inventors. If AI-generated discoveries cannot be protected, capital will not fund decade-long trials, and well-funded labs may hide results as trade secrets. One path is to count problem-setting and constraint design as human inventorship. Another is a new licensing wrapper for machine-generated findings.
There is no original experiment in this paper. The quantitative claims are citations used to size the validation bottleneck.
| Claim | Number | Source |
| Capitalised cost of a new molecular entity | about $2.6 billion | DiMasi et al., 2016 |
| Inflation-adjusted FDA drug development cost | roughly doubles every 9 years | Eroom's Law, Scannell et al., 2012 |
| Computationally predicted crystal structures | millions | Merchant et al., 2023 |
| Data access tiers | 4 | Table 1 |
| Hypothesis exchange phases | 4 | Table 2 |
The baselines are today's institutions: peer review, open data, CROs, FDA priority-review vouchers, vaccine AMCs. Nothing here is simulated, and nothing is shown to clear a market.
The paper's sharpest empirical prediction is jagged acceleration. Mathematics and theoretical computer science can drive validation cost down with tools such as Lean. Clinical trials, multi-environment crop phenotyping, and collider time are bound to biological clocks and mega-facilities. An AI scientist cannot shorten those clocks. A physics and chemistry simulation benchmark (Cao et al., 2026) is cited to show that agents which ignore execution cost can be far more expensive than classical parameter sweeps.
The useful reframe for anyone shipping an AI scientist: more reasoning without cost-awareness just lengthens the lab queue. The missing layer is a protocol for scarce resources across institutions.
The full on-chain science economy is a blueprint, not a product. Near-term pieces are more concrete. Stage hypotheses through computation before wet work. Log negative results. Score agents on token cost and logistical feasibility, not accuracy alone.
This is incremental institutional design, not a new training method. It recasts AI-for-science as a multi-agent allocation problem. The disclaimer is unusually plain: these are the authors' personal views, not Google DeepMind policy.
The authors draw their own capability line. Current AI scientists are good at recombinant, Kuhn-normal work and weak at paradigm shifts. Homogeneous training data risks a cognitive monoculture, with many agents charging the same wall. An AI-assisted attack on a Navier-Stokes Millennium problem prompted a group of mathematicians to sign "A Severe Misalignment of AI in Mathematics", arguing that automating proofs can work against human understanding. Uncritical human oversight produces an illusion of understanding.
The design is also loose. Smart contracts and IP-NFTs sit next to a permissioned clearinghouse modeled on SWIFT, because dual-use data may not belong on a public chain. Shapley-style data prices and exact marginal credit are admitted to be intractable. Epistemic staking favors well-capitalized labs. Persistent identity for copyable agents, Sybil resistance, and collusion detection remain open.
This is a cs.CY position paper. No ablation, no pilot, no failure case. Eroom's Law and the $2.6 billion figure come from older drug-development literature. Read it as a design memo. Treating it as a deployable system overstates readiness. Dual-use discussion stays at compute governance and structured access; screening thresholds are not specified.