Hypothesis agents

The task

BeeARD’s hackathon brief: given a subgraph — a short path an explorer agent has walked through a knowledge graph of biomedical concepts, say Inflammation → increases → Amyloid Beta → accumulates in → Alzheimer's Disease — build a multi-agent system that returns a novel, defensible research hypothesis. The subgraph is the creative seed; the agents have to turn it into a claim worth testing.

What we built

A five-agent pipeline on LangGraph, with a critic at the end that can send the whole thing back around:

1. Graph analyst — reads the subgraph and produces a structured mechanistic summary: the key entities, the relationships between them, and where the path is thin.
2. Context agent — enriches each entity with what it actually is, querying UniProt, Ensembl and DisGeNET rather than relying on the model’s memory.
3. Evidence agent — runs in parallel with the context agent, pulling supporting and contradicting literature from PubMed, Europe PMC and bioRxiv.
4. Hypothesis generator — synthesises the summary, context and evidence into a stated hypothesis with references.
5. Critic analyst — evaluates it and returns ACCEPT or a critique. On a critique the state loops back to the graph analyst, up to three times.

The retrieval agents run on a small, fast model; the analyst, generator and critic run on a reasoning model, since those are the steps where the quality of the thinking sets the ceiling.

Why the grounding matters

Grounding every entity in a real database and every claim in retrieved literature is what gives the critic something to weigh. It is evaluating a hypothesis against evidence, not judging prose — and that is the difference between a loop that converges and one that just rephrases.