Agentic Economies for Autonomous Scientific Discovery

Nenad Tomasev, Matija Franklin, Atoosa Kasirzadeh, Vivek Natarajan, Alan Karthikesalingam, Sebastien Krier, Simon OsinderoarXiv 2026 · arXiv

TL;DR

This paper argues that the next frontier for AI in science isn't just better reasoning, but better resource management. The authors propose a formal "Agentic Economy"—a market-based infrastructure using smart contracts and decentralized science (DeSci) primitives—to coordinate multi-agent AI systems, prioritize expensive physical experiments, and ensure equitable, safe discovery.

What Problem Does This Paper Address?

While AI has made massive strides in "cognitive" tasks like protein folding (AlphaFold) and hypothesis generation, scientific discovery is ultimately limited by the physical world. Experiments are expensive, time-consuming, and resource-constrained—a phenomenon exemplified by "Eroom’s Law" in drug discovery, where R&D costs double every nine years despite better technology.

The authors identify a looming "validation bottleneck": if AI agents can generate millions of plausible hypotheses for free, they will quickly overwhelm our laboratories. Currently, we lack a system to coordinate these autonomous agents across different institutions, price their contributions, manage limited lab time, or handle the legal and safety liabilities of AI-driven breakthroughs.

Methods

The paper is a positional framework rather than an empirical study. It synthesizes concepts from multi-agent systems, market design, and decentralized finance to propose a "Hypothesis Exchange Pipeline." Key components include:

  • Scientific Data Markets: A tiered system for data access (from Open Access to Confidential) using Trusted Execution Environments (TEEs) and Zero-Knowledge Proofs to allow agents to query sensitive data without compromising privacy.
  • Idea Tokenization: Using IP-NFTs (Intellectual Property Non-Fungible Tokens) to turn hypotheses into tradable epistemic assets.
  • Automated Coordination: Utilizing Smart Contracts to act as "computable Data Use Agreements" (cDUAs), allowing agents to negotiate access to compute and lab resources autonomously.
  • Reputation Systems: Implementing cryptographic staking and audit trails to prevent "data farming" (generating junk experiments for fees) and collusion.

Datasets

As a framework paper, it does not introduce a new training dataset. However, it references several emerging benchmarks designed to evaluate "cost-aware" AI agents, such as:

  • Simulcost: A benchmark for automating physics simulations with an eye on token costs.
  • Ainstein-bench: For benchmarking coding agents on scientific repositories.
  • Scicode: A research coding benchmark curated by scientists.

Key Insights and Ideas

  • The "Oracle Problem": A major hurdle is bridging the digital ledger (where AI lives) and the physical lab. The authors suggest "hardware roots of trust" at the sensor level to provide immutable proof that an experiment actually happened.
  • Valuing Negative Results: AI can share failures much more efficiently than humans, who often hide them due to publication bias. The paper proposes "Negative Result Registries" where agents are compensated for pruning "dead-end" research paths.
  • Agent-Native Research Artefacts: The authors argue that the PDF is an "engineering tax" on AI. They propose moving toward machine-executable "knowledge packages" that include the cognitive reasoning trace, the physical execution code, and the raw data provenance.
  • Algorithmic Cross-Subsidization: To prevent AI from only pursuing "profitable" science (like drugs for wealthy demographics), the authors suggest smart contracts that automatically route a micro-percentage of commercial licensing fees into a treasury for basic, public-good research.

Key Results

The paper provides a conceptual roadmap rather than quantitative benchmarks. Its primary "results" are two structural models:

  1. A Four-Tier Data Access Model: Ranging from public APIs (Tier 1) to secure multi-party computation for sensitive genomic/EHR data (Tier 4).
  2. The AI Idea Lifecycle: A four-stage process (Anchoring → Ex-Ante Evaluation → Brokerage → Validation Payout) that separates the cognitive labor of discovery from the physical labor of execution.

Limitations and Future Work

  • Cognitive Monoculture: If all AI scientists are trained on the same data, they may all chase the same ideas, leading to a lack of diversity in research.
  • Market Manipulation: The system must be resilient against "Sybil attacks," where a malicious actor creates thousands of fake agents to artificially inflate the importance of a specific research direction.
  • The Human Element: There is a risk that over-reliance on AI could erode human expertise and creativity. The authors emphasize that humans must shift from "execution" to "orchestration and oversight."
  • Dual-Use Risks: Autonomous labs could be hijacked to synthesize pathogens or toxins. The authors call for "agentic overseers" to screen all synthesis requests against predictive toxicity models.

Summary updated

Abstract

Recent advances in agentic Artificial Intelligence (AI) systems have marked a shift in AI for Science: moving away from the use of individual AI systems for narrow task execution, toward multi-agent systems capable of orchestrating complex, end-to-end research workflows and performing (semi-)autonomous scientific discovery. The development of multi-agent AI-for-science systems has primarily focused on improving the cognitive capabilities of AI systems, specifically by making advanced reasoning and hypothesis generation more reliable. However, focusing only on cognitive capability improvement could ignore appropriate management of resources, a key bottleneck in scientific discovery. Testing and validating scientific hypotheses and experiments is, physically and economically, resource-intensive and resources are limited. This means that to make significant advancements in autonomous scientific discovery, such as improving human-AI co-scientist complementarity or reaching a truly closed-loop automated process, we must pair ongoing improvements in reasoning capabilities with robust resource management. In this paper, we outline an infrastructure for AI resource management by developing the necessary foundations of scientific agent economies, markets, and institutions. The aim of this infrastructure is to empower AI agents and human scientists to effectively (i) collaborate and establish research priorities, (ii) assign credit, (iii) track accountability and liability, and (iv) safeguard against malicious use and information security risks. Finally, we engage with the macro-level societal implications. of (semi-)autonomous scientific discovery to inform the development of governance policies ensuring an equitable distribution of AI-driven discoveries and derivative future technologies.
BibTeX
@misc{tomasev2026agentic,
  author = {Nenad Tomasev and Matija Franklin and Atoosa Kasirzadeh and Vivek Natarajan and Alan Karthikesalingam and Sebastien Krier and Simon Osindero},
  title = {{Agentic Economies for Autonomous Scientific Discovery}},
  year = {2026},
  eprint = {2609.31562},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2609.31562},
}
Figure