Lab Newsletter — June 27, 2026: Agents Get Put to the Test

A theme today: the field is putting agents through their paces — in the literature, at the microscope, and against real reference data. Here’s what caught our eye.

🤖 “AI scientists” still flunk the library

A sobering new benchmark, AutoResearchBench, tests whether AI agents can do the unglamorous first step of research — finding the right papers. Two tasks: “deep research” (track down a specific target paper through multi-step probing) and “wide research” (collect every paper matching a set of conditions). Even the strongest LLM agents manage only ~9% (9.39% accuracy / 9.31% IoU), with most baselines below 5% — despite many having “conquered” general web-browsing benchmarks. Why it matters for the lab: our autonomous-research-agents live exactly here. The takeaway isn’t “agents can’t” — it’s that literature discovery is a real, unsolved bottleneck, and human-in-the-loop checks (and adversarial cross-verification, as some new multi-agent research frameworks propose) are well-placed bets, not training wheels.

🔬 Agents come to ImageJ — with reproducibility built in

Agentic-J (Johanns et al., arXiv, June 2026) is a containerized, multi-agent assistant for Fiji/ImageJ: a biologist asks in plain language (“segment the nuclei, track the cells, quantify per condition”) and specialized sub-agents handle plugin selection, code generation, debugging, QA and statistical reporting — writing every decision into a documented, reproducible project. It ships a full Fiji distribution in Docker, keeps the familiar interface (human-in-the-loop, not black box), and talks to napari over the Model Context Protocol. Why it matters for the lab: this is precisely the pattern we build toward with ImJoy, ImageJ.JS and the BioImage.IO Chatbot — agents wrapped around trusted tools, reproducible by construction, and speaking MCP like our own stack.

🧬 The Human Cell Atlas convenes; the retina gets mapped

The Human Cell Atlas General Meeting (Boston, June 16–18) gathered the global single-cell community to push shared standards for data and spatial biology — alongside a new collaboration widening access to single-cell multiomics. In the same spirit, a Human Retina Cell Atlas integrates ~3.9M cells from 125 donors into 130+ cell types and ties them to GWAS/eQTL signals. Why it matters for the lab: standardized reference atlases are the substrate the virtual cell — and our image × omics models — learn from; the boring work of standards is what makes the exciting models trustworthy.

📖 From the lab

A quiet point of pride: our own BioImage.IO Chatbot, supported by SciLifeLab’s DDLS program, keeps growing from a Q&A helper into a full agent that reads papers, drafts experimental plans, and drives microscopes and liquid handlers — the same agent-meets-instrument direction this whole issue circles around.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. Have lab news to share — a talk, paper, conference or release? Message me on Slack.