6 min read

Agentic pipelines land in drug discovery

Agentic pipelines land in drug discovery
Nº 01 · The Lede X Computational biology

Muni Bio runs autonomous discovery pipeline

Muni Bio runs autonomous discovery pipeline
Fig. IX · Filed 28 Jul 2026.

Muni Bio ran an early-stage drug discovery pipeline designed and executed autonomously by agents, walking through target selection, hit generation, and prioritization end-to-end in a public writeup. The lab is betting the default shape of early discovery flips from human-driven with AI assists to agent-driven with human review. It's one of the first end-to-end demonstrations from an operating drug-discovery outfit, not a benchmark-only paper.

Read the source

The physical layer of autonomous science
Fig. IIX · Filed 28 Jul 2026.
Nº 02 X Field report

The physical layer of autonomous science

Autonomous discovery's bottleneck isn't the model, argues a widely-shared thread — it's the physical layer connecting agents to instruments, reagents, and robotics. The frame reframes the competitive frontier away from pure LLM benchmarks toward wet-lab integration, and lines up the automation-hardware stack as the next contested budget item in bio-AI — the same argument DeepMind made last week.

Read more
Plato-Bio benchmarks novelty claims
Fig. IIIarXiv · Filed 28 Jul 2026.
Nº 03 arXiv Benchmarks · Evaluation

Plato-Bio benchmarks novelty claims

Plato-Bio verifies biological novelty by rediscovering known findings from time-sliced literature, then scoring structural plausibility of new proposals. The setup turns "our agent discovered something new" from a marketing claim into a checkable one, and anchors a reference benchmark for the flood of autonomous-discovery papers now landing weekly.

Read more
Also Filed · Five Briefs from the queue
Nº 04 arXiv Agents · Infrastructure

Agent debugs ECG classifiers by failure

An evidence-driven agent recursively refines ECG classifiers by examining what the model gets wrong, not just its aggregate metrics. Moves clinical-ML iteration from human-audited error analysis to an agent-driven loop — narrows the gap between research-grade classifiers and the failure-mode discipline clinical deployment actually requires.

Read
Nº 05 bioRxiv Spatial transcriptomics

STAR Suite ships single-binary transcriptomics

STAR Suite bundles reproducible transcriptomics processing into a single executable designed for AI-agent orchestration. Raises the floor for what "agent-ready" bio tooling looks like — no environment hell, no dependency drift, one binary an agent can call.

Read
Nº 06 bioRxiv Computational biology

MetaClaw audits metagenomic analysis

MetaClaw runs metagenomic and multi-omics analysis end-to-end with a built-in audit trail, logging every decision the agent makes. Provenance-by-default in an omics agent — moves auditability from a compliance afterthought to a design constraint, which is where any clinical or regulatory adoption starts.

Read
Nº 07 Anthropic Field report

Anthropic stakes open-weights position

Dario Amodei laid out Anthropic's position on open-weights models, framing when release is safe and when it isn't. Sharpens the policy debate the field will inherit as open-weight biology models keep shipping from Chinese and academic labs.

Read
Nº 08 Anthropic Field report

Claude Opus 5 lands

Anthropic shipped Claude Opus 5 , pitched as a step change for long-running agents plus coding and professional work — the tier most bio agent builders default to when reliability across many tool calls matters more than latency.

Read

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

Agentic Discovery  ·  Nº 64  ·  28 Jul 2026

Editor's Note

Tuesday brings the shift we've been circling for a year: agentic workflows moving from demo to default, with the infrastructure debate right behind.

 

Nº 01 · The Lede  —  X  —  Computational biology

Muni Bio runs autonomous discovery pipeline

Muni Bio runs autonomous discovery pipeline

Fig. I  X · Filed 28 Jul 2026.

Muni Bio ran an early-stage drug discovery pipeline designed and executed autonomously by agents, walking through target selection, hit generation, and prioritization end-to-end in a public writeup. The lab is betting the default shape of early discovery flips from human-driven with AI assists to agent-driven with human review. It's one of the first end-to-end demonstrations from an operating drug-discovery outfit, not a benchmark-only paper.

Read the source →

Why it matters

Moves agentic pipelines from proof-of-concept to a stated operating thesis at a working discovery lab — the reference point for whether "autonomous discovery" is a slide or a workflow just moved from academia to industry.

 

Nº 02  —  X  —  Field report

The physical layer of autonomous science

Fig. II  X · Filed 28 Jul 2026.

The physical layer of autonomous science

Autonomous discovery's bottleneck isn't the model, argues a widely-shared thread — it's the physical layer connecting agents to instruments, reagents, and robotics. The frame reframes the competitive frontier away from pure LLM benchmarks toward wet-lab integration, and lines up the automation-hardware stack as the next contested budget item in bio-AI — the same argument DeepMind made last week.

Read more →

 

Nº 03  —  arXiv  —  Benchmarks · Evaluation

Plato-Bio benchmarks novelty claims

Fig. III  arXiv · Filed 28 Jul 2026.

Plato-Bio benchmarks novelty claims

Plato-Bio verifies biological novelty by rediscovering known findings from time-sliced literature, then scoring structural plausibility of new proposals. The setup turns "our agent discovered something new" from a marketing claim into a checkable one, and anchors a reference benchmark for the flood of autonomous-discovery papers now landing weekly.

Read more →

 

Also Filed  ·  Five Briefs from the queue

Nº 04  —  arXiv  —  Agents · Infrastructure

Agent debugs ECG classifiers by failure

An evidence-driven agent recursively refines ECG classifiers by examining what the model gets wrong, not just its aggregate metrics. Moves clinical-ML iteration from human-audited error analysis to an agent-driven loop — narrows the gap between research-grade classifiers and the failure-mode discipline clinical deployment actually requires.

Read →

Nº 05  —  bioRxiv  —  Spatial transcriptomics

STAR Suite ships single-binary transcriptomics

STAR Suite bundles reproducible transcriptomics processing into a single executable designed for AI-agent orchestration. Raises the floor for what "agent-ready" bio tooling looks like — no environment hell, no dependency drift, one binary an agent can call.

Read →

Nº 06  —  bioRxiv  —  Computational biology

MetaClaw audits metagenomic analysis

MetaClaw runs metagenomic and multi-omics analysis end-to-end with a built-in audit trail, logging every decision the agent makes. Provenance-by-default in an omics agent — moves auditability from a compliance afterthought to a design constraint, which is where any clinical or regulatory adoption starts.

Read →

Nº 07  —  Anthropic  —  Field report

Anthropic stakes open-weights position

Dario Amodei laid out Anthropic's position on open-weights models, framing when release is safe and when it isn't. Sharpens the policy debate the field will inherit as open-weight biology models keep shipping from Chinese and academic labs.

Read →

Nº 08  —  Anthropic  —  Field report

Claude Opus 5 lands

Anthropic shipped Claude Opus 5 , pitched as a step change for long-running agents plus coding and professional work — the tier most bio agent builders default to when reliability across many tool calls matters more than latency.

Read →

 

· · ·

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