Agents get a wire protocol

Agentic Discovery  ·  Nº 29  ·  03 Jun 2026

Editor's Note

Today: a serious bid to standardize how agents talk to lab instruments, plus protein dark matter starts giving up its secrets.

Nº 01 · The Lede  —  arXiv  —  Agents · Infrastructure

LAP standardizes agent-instrument talk

Fig. I  arXiv · Filed 03 Jun 2026.

LAP proposes a wire protocol — Lab Agent Protocol — for autonomous agents to drive lab instruments directly, in the same spirit MCP (Model Context Protocol, Anthropic's spec for agents-to-tools) standardized software integrations. The arxiv preprint sketches discovery, capability negotiation, and safety interlocks so an LLM agent can address a plate reader or liquid handler without bespoke glue code. Early demos cover pipetting and microscopy.

Read the source →

Nº 02  —  bioRxiv  —  Structural biology · Protein design

Fig. II  bioRxiv · Filed 03 Jun 2026.

Protein dark matter, illuminated

Embedding-based clustering cracks functional dark matter — proteins with no annotated function — by routing ESM-style embeddings through downstream ML classifiers to surface targeted function calls. The bioRxiv preprint reports new functional assignments in previously opaque protein families, moving protein language models from generic feature extractors to direct annotation engines and shrinking the unannotated fraction of UniProt that everyone has learned to live with.

Read more →

Nº 03  —  arXiv  —  Field report

Fig. III  arXiv · Filed 03 Jun 2026.

Grading chemistry reasoning step-by-step

Process-level evaluation arrives for chemical reasoning in LLMs: instead of scoring final answers, the benchmark verifies intermediate states across a synthesis or mechanism trace. Anchors a new floor for chemistry-applicable model claims — a model can no longer guess its way to a right answer and call it reasoning, which matters anywhere agents are proposing routes for medicinal chemistry.

Read more →

Also Filed  ·  Two Briefs from the queue

Nº 04  —  bioRxiv  —  Field report

Tabular foundation models close ancestry gap

Polygenic risk scores trained with tabular foundation models transfer across ancestries far better than standard GWAS-derived PRS, per a new bioRxiv preprint. Narrows one of the most stubborn equity gaps in genomic medicine, where models trained on European cohorts have routinely underperformed elsewhere.

Read →

Nº 05  —  Anthropic  —  Field report

Claude Opus 4.8 ships

Anthropic released Claude Opus 4.8, its top-tier model, with gains on coding, agentic tasks, and long-running work consistency. Raises the floor for what counts as a viable backbone for multi-hour biology agent runs — the failure mode that has kept most autonomous discovery demos under an hour.

Read →

· · ·

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