The experiment
What if the next scientific breakthrough is already published — just scattered across papers that no single researcher reads together?
What MAGELLAN is
MAGELLAN (Multi-Agent Generative Exploration of Latent Links Across kNowledge) builds on a simple, proven idea: important scientific connections already exist in the published literature, but go unnoticed because no one reads across all the relevant fields. Don Swanson proved this in 1986. We scale it with AI.
The system deploys 12 specialized AI agents that read across disciplinary boundaries, identify where established knowledge in one field connects to another through an unexplored bridge, and generate testable hypotheses about what that connection means. Then it attacks its own ideas — adversarially, rigorously — and presents only the survivors.
An honest framing
This is an experiment. We don't know yet if it works.
We believe multi-agent AI can find connections humans miss — but every hypothesis here needs testing by real scientists. We're not claiming discoveries. We're claiming a system that generates hypotheses worth investigating.
The confidence scores are intentionally moderate (4-5/10). The kill rate is intentionally high (86%). We tag every claim as GROUNDED, PARAMETRIC, or SPECULATIVE.
We label our uncertainty because that's what makes this credible. If even one hypothesis leads to a validated experiment, MAGELLAN will have done something meaningful. But the honest answer is: we don't know yet. That's why we need you.
Growing with AI
MAGELLAN is built on March 2026 frontier models — Claude Opus 4.6 and Sonnet 4.6. The architecture is designed so that as models improve, every stage of the pipeline becomes more powerful:
Session 4 is the beginning. We're building the infrastructure — the agents, the quality gates, the meta-learning loops — so that when models 10x more capable arrive, the system is ready.
Open source
The full pipeline — every agent prompt, every scoring rubric, every quality gate — is open source. We believe scientific discovery tools should be transparent and reproducible. If you can improve a scoring rubric or add a new agent strategy, we want your contribution.
View on GitHub →Contact
Questions, collaborations, or feedback: hello@magellan-discover.ai