ScienceClaw A hackathon for building decentralized collectives of AI agents that solve real scientific and technical problems.
Purpose
ScienceClaw explores when decentralized collective intelligence can produce capabilities beyond a single agent or conventional workflow.
Teams build working collectives of specialized agents, models, simulations, robots, sensors, and laboratory tools. Different members of the collective can hold distinct knowledge and capabilities, coordinate and adapt, and leave traceable scientific artifacts, provenance, findings, and unmet needs that the collective can build upon—without requiring one central planner to prescribe the entire process.
Challenge
Teams identify a difficult scientific or technical problem and build a functioning agentic system to address it. Each project should demonstrate:
Problem
Identify a difficult, open-ended, consequential scientific or technical challenge.
Functioning System
Build a working agentic system — not just a concept or architecture.
Result
Demonstrate what the system actually accomplishes.
Validation
Provide convincing evidence, benchmarks, or real-world validation.
The goal is not simply to propose an idea or agent architecture, but to demonstrate a working system and what it can accomplish.
Resources & Technology
ScienceClaw is model- and framework-agnostic. Teams can combine frontier models, open models, their own models, existing agent frameworks, scientific datasets, compute, robots, sensors, and laboratory infrastructure.
Scientific datasets
Shared corpora and domain data for grounding investigations.
AI models & tools
Foundation models, agent frameworks, and open tooling to compose into collectives.
Compute & tokens
A common pool of credits so experimentation is never bottlenecked.
Hardware
Robots, microcontrollers, and sensors for agents that act in the world.
Lab access
Cloud labs and experimental infrastructure for closing the loop on real experiments.
Open problem space
Sponsor-led challenges provide concrete inspiration — unexpected approaches welcome.
The goal is not to prove that one model is best. Teams can use, reuse, modify, and combine available technology—including capabilities developed during the event—to build a functioning decentralized system that accomplishes something scientifically meaningful.
Challenge Areas
Scientific domains are intentionally open. We provide challenge seeds, problems, datasets, resources, or sponsor-led challenges to provide entry points without constraining what teams build.
Biology & Protein Design
Collectives that hypothesize, design, and iterate across sequence, structure, and function.
Robotics & Physical Agency
Agents that reach beyond the browser into robots, sensors, and cloud laboratories.
Manufacturing & Materials
From candidate discovery to fabrication — closing the loop between simulation and matter.
Open Scientific Challenge
Participant-defined problems. Bring the question you can't stop thinking about.
Teams can formulate their own problems and work across challenge areas. There is one overall competition, rather than separate tracks.
Judging
Science is the benchmark. Success is not simply whether an agent completed a task, but whether the system produced a scientifically meaningful result supported by evidence, validation, or measurable progress.
The judging rubric also asks whether organizing capabilities as a decentralized collective adds something beyond a single agent, isolated agents, or a conventional workflow.
Problem Significance & Complexity
How difficult, open-ended, consequential, and scientifically meaningful is the problem? Does solving it meaningfully advance a scientific or technical goal?
Scientific/Technical Impact
What was actually achieved? Is the result novel, meaningful, rigorous, and potentially useful?
Decentralized Agency
Are decision-making and capabilities meaningfully distributed across agents? Do agents have distinct information, resources, or capabilities? Can they coordinate and adapt without a central planner prescribing the entire trajectory?
Collective Capability
Does organizing agents collectively add measurable capability? Where possible, demonstrate this through a simple comparison such as multi-agent vs. single-agent, collaborative vs. isolated, or decentralized vs. centralized.
Execution and Validation
Is there a functioning system rather than just a concept? Is its performance demonstrated with convincing evidence, benchmarks, or real-world validation?
Collaboration Bonus: +10 — Collaboration is not part of the core judging score. Teams can earn up to +10 points by creating capabilities that materially strengthen other teams, such as:
- A tool or agent successfully reused by another team
- Data, discoveries, or experiments that enable another team's result
- Infrastructure that improves multiple systems
- Capabilities that propagate through the ecosystem
The Hackathon as a Multi-Agent System
The hackathon operates across multiple levels:
Agents
Collaborate within teams to solve problems.
Teams
Exchange capabilities across teams.
Ecosystem
Successful capabilities propagate because they increase the effectiveness of other systems.
The competition therefore becomes a live experiment in decentralized collective intelligence: teams compete to build strong systems while also having an incentive to make the broader ecosystem stronger.
Join the collective.
October 30 – November 1, 2026. Bring a problem — or come find one.
Apply now