Agents stop assisting and start discovering

Agentic Discovery  ·  Nº 118  ·  12 Oct 2026

Editor's Note

Monday, and the week opens with agents that didn't summarize the literature so much as add to it.

Nº 01 · The Lede  —  bioRxiv  —  Agents · Infrastructure

Autonomous agents find new reverse transcriptase enzymes

Fig. I  bioRxiv · Filed 12 Oct 2026.

Autonomous agents discovered new reverse transcriptases carrying tandem repeat arrays, according to a bioRxiv preprint from P. H. Yoon and colleagues. The agents ran the search themselves rather than handing a ranked candidate list to a human for triage. Reverse transcriptases are the engine behind prime editing and other retroelement-derived tools, so a new structural variant is a find in its own right, separate from how it was found. Most agent-in-biology work so far has demonstrated workflow automation: faster annotation, faster literature synthesis, faster code. This one reports a discovery, which moves research agents from software that compresses existing work to instruments that return novel biological entities.

Read the source →

Nº 02  —  bioRxiv  —  Field report

Fig. II  bioRxiv · Filed 12 Oct 2026.

CHORUS maps membrane binding sites from sequence alone

CHORUS maps membrane interaction interfaces across the proteome using embeddings pulled from multiple protein language models — numerical representations a model learns from amino-acid sequence, with no structure supplied. Membrane-binding surfaces have been among the harder annotations to get at scale, since the contacts are often transient and partly disordered. Delivering them at proteome scale puts peripheral membrane binding on the same footing as the structural annotation AlphaFold normalized, and makes a sequence-only first pass viable where a structure pipeline was previously the only route.

Read more →

Nº 03  —  arXiv  —  Agents · Infrastructure

Fig. III  arXiv · Filed 12 Oct 2026.

SAIL puts a science-aware loop inside research agents

SAIL proposes a science-aware loop as the organizing idea for research agents: put the shape of scientific work into the agent's control cycle rather than borrowing the generic plan-act-observe scaffolding that came out of software engineering tasks. The arXiv paper, credited to the SAIL Model Team, treats the loop itself as the thing to design, not only the language model running inside it. That shifts where scientific-agent quality is argued to come from, and opens a line of competition that doesn't require frontier-scale compute to enter.

Read more →

Also Filed  ·  One Brief from the queue

Nº 04  —  arXiv  —  Clinical AI · Evaluation

LLMs flag spin in clinical trial write-ups

LLMs detect spin in clinical trial reports, the practice of writing up a study so a null primary endpoint still reads as a win. Automating that check turns overstatement into a measurable property of the trial literature rather than a complaint about it.

Read →

· · ·

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