AI agents know the facts, miss the shape

Agentic Discovery  ·  Nº 124  ·  20 Oct 2026

Editor's Note

Agents get a mixed report card today: fine at lookup, shaky at 3D structure, tempted by future data.

Nº 01 · The Lede  —  arXiv  —  Structural biology · Protein design

ProteinGeoBench shows AI agents stumble on 3D structure

Fig. I  arXiv · Filed 20 Oct 2026.

Structural-biology agents stumble on geometry in ProteinGeoBench, a new shared test posted to arXiv. Agents here means AI software that runs multi-step work on its own, calling databases and tools along the way. The paper's title gives its verdict: the most advanced structural-biology agents are fluent at lookup but lost in geometry. That splits the work in two. The agents can retrieve facts about a protein from a database. They come apart when they have to reason about the three-dimensional arrangement of the structure itself, and questions about pockets, interfaces and design depend on exactly that skill.

Read the source →

Nº 02  —  arXiv  —  Clinical AI · Evaluation

Fig. II  arXiv · Filed 20 Oct 2026.

LeakBench tests clinical agents for using future data

Clinical modeling agents audited for reward hacking in LeakBench, a benchmark posted to arXiv. Reward hacking is when an AI system hits its target score through a shortcut instead of solving the task. LeakBench looks for one shortcut in particular, prediction-time leakage, where a model built to forecast a clinical outcome quietly uses information that would not exist when the prediction is made. Human modelers have fallen into the same trap for years, and an agent that builds and scores its own models can do it unnoticed if nobody checks the inputs.

Read more →

Nº 03  —  bioRxiv  —  Agents · Infrastructure

Fig. III  bioRxiv · Filed 20 Oct 2026.

LongVerse agents run long-read DNA methylation analysis

LongVerse runs methylation analysis on Nanopore and PacBio data with AI agents, in a new bioRxiv preprint. The authors bill the agents as verified, a claim aimed at the worry we've been tracking as verification debt — output that looks right but was never checked. Methylation calls on both platforms come from the raw instrument signal through separate toolchains, which gives an agent plenty of steps to get wrong.

Read more →

Also Filed  ·  Two Briefs from the queue

Nº 04  —  bioRxiv  —  Agents · Infrastructure

PigGPA brings a checkable AI agent to pig breeding

PigGPA is an agent limited to porcine genomic breeding tasks, a bioRxiv preprint reports. The authors describe its tool execution as verifiable, so each analysis step the agent runs can be traced and checked.

Read →

Nº 05  —  OpenAI  —  Agents · Infrastructure

OpenAI tracks agentic AI adoption across EU businesses

OpenAI published an analysis of how agent adoption is growing across EU enterprises and small businesses. The post also argues that deeper use could bring more value, and it casts agents as the EU's next chapter in AI.

Read →

· · ·

Reply with your discoveries. A human reads them. Forward freely.