5 min read

Agents creep into real lab work

Agents creep into real lab work
Nº 01 · The Lede Axios Field report

Codex usage data lands

Codex usage data lands
Fig. IAxios · Filed 30 Jun 2026.

Codex agent usage is now outrunning ChatGPT inside OpenAI by token volume, per a joint report with Columbia, Duke, and Penn — and outside organizations are catching up fast. The study breaks adoption into OpenAI staff, external orgs, and individuals, and the trajectories converge faster than prior agent rollouts. Pairs with OpenAI's own X thread showing agents now running cross-functional work in every department. First credible third-party measurement of where delegated work actually sits on the adoption curve.

Read the source

Tahoe's cell-perturbation platform ships
Fig. IIX · Filed 30 Jun 2026.
Nº 02 X Cell biology · Funding

Tahoe's cell-perturbation platform ships

Tahoe AI moved from vision to product this week, with founder commentary from genophoria highlighting years of build toward large-scale cell-perturbation data for AI training. Tahoe's pitch — industrial-scale perturb-seq paired with AI models — turns the 100M-cell dataset class into a usable substrate. Moves virtual-cell modeling one step closer to having training data that matches its ambitions.

Read more
Diffusion targets PPI interfaces
Fig. IIIbioRxiv · Filed 30 Jun 2026.
Nº 03 bioRxiv Field report

Diffusion targets PPI interfaces

Pep2Mol generates 3D molecules aimed directly at protein-protein interfaces, using diffusion models conditioned on the peptide-binding pocket. PPIs have been the canonical "undruggable" target class for two decades; a generative model that natively respects interface geometry shifts what the small-molecule frontier looks like for this class.

Read more
Also Filed · Three Briefs from the queue
Nº 04 arXiv Field report

LLM diagnosis: competent, inconsistent

Clinical Reasoning Graphs evaluate LLM diagnostic reasoning structurally — and find models reach the right answer but take wildly different paths run to run. Establishes reasoning-trace consistency, not just final-answer accuracy, as the bar for clinical-grade LLM deployment — the same process-over-answer standard BiomniBench introduced for biomedical agents.

Read
Nº 05 bioRxiv Field report

Germinal pipeline goes open-source

OpenGerminal reimplements the Germinal antibody design pipeline as open source. Closes the reproducibility gap that has shadowed proprietary antibody-generation stacks and gives the field a shared baseline for benchmarking new designers.

Read
Nº 06 arXiv Drug discovery · Computational

Molecular optimization beyond drugs

The NMO benchmark extends molecular optimization evaluation past drug-like chemistry into nanotechnology design space. Broadens what generative chemistry models are graded on — drug-likeness was never the only optimization target that matters.

Read

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

Agentic Discovery  ·  Nº 46  ·  30 Jun 2026

Editor's Note

Tuesday brief: Codex usage data finally lands, Tahoe ships, and three preprints worth your scroll.

 

Nº 01 · The Lede  —  Axios  —  Field report

Codex usage data lands

Codex usage data lands

Fig. I  Axios · Filed 30 Jun 2026.

Codex agent usage is now outrunning ChatGPT inside OpenAI by token volume, per a joint report with Columbia, Duke, and Penn — and outside organizations are catching up fast. The study breaks adoption into OpenAI staff, external orgs, and individuals, and the trajectories converge faster than prior agent rollouts. Pairs with OpenAI's own X thread showing agents now running cross-functional work in every department. First credible third-party measurement of where delegated work actually sits on the adoption curve.

Read the source →

Why it matters

Anchors a real number under the "agents are replacing knowledge work" claim — biology orgs evaluating Codex-style platforms now have a usage curve to benchmark against instead of vendor decks.

 

Nº 02  —  X  —  Cell biology · Funding

Tahoe's cell-perturbation platform ships

Fig. II  X · Filed 30 Jun 2026.

Tahoe's cell-perturbation platform ships

Tahoe AI moved from vision to product this week, with founder commentary from genophoria highlighting years of build toward large-scale cell-perturbation data for AI training. Tahoe's pitch — industrial-scale perturb-seq paired with AI models — turns the 100M-cell dataset class into a usable substrate. Moves virtual-cell modeling one step closer to having training data that matches its ambitions.

Read more →

 

Nº 03  —  bioRxiv  —  Field report

Diffusion targets PPI interfaces

Fig. III  bioRxiv · Filed 30 Jun 2026.

Diffusion targets PPI interfaces

Pep2Mol generates 3D molecules aimed directly at protein-protein interfaces, using diffusion models conditioned on the peptide-binding pocket. PPIs have been the canonical "undruggable" target class for two decades; a generative model that natively respects interface geometry shifts what the small-molecule frontier looks like for this class.

Read more →

 

Also Filed  ·  Three Briefs from the queue

Nº 04  —  arXiv  —  Field report

LLM diagnosis: competent, inconsistent

Clinical Reasoning Graphs evaluate LLM diagnostic reasoning structurally — and find models reach the right answer but take wildly different paths run to run. Establishes reasoning-trace consistency, not just final-answer accuracy, as the bar for clinical-grade LLM deployment — the same process-over-answer standard BiomniBench introduced for biomedical agents.

Read →

Nº 05  —  bioRxiv  —  Field report

Germinal pipeline goes open-source

OpenGerminal reimplements the Germinal antibody design pipeline as open source. Closes the reproducibility gap that has shadowed proprietary antibody-generation stacks and gives the field a shared baseline for benchmarking new designers.

Read →

Nº 06  —  arXiv  —  Drug discovery · Computational

Molecular optimization beyond drugs

The NMO benchmark extends molecular optimization evaluation past drug-like chemistry into nanotechnology design space. Broadens what generative chemistry models are graded on — drug-likeness was never the only optimization target that matters.

Read →

 

· · ·

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