Lab Newsletter — June 25, 2026: Agents Grow Up, Virtual Cells Get a Reality Check

A quieter week on the model front, a louder one on what we build around the models. Here’s what caught our eye — with a strategy-radar lens on where the lab is heading.

🤖 Anthropic bets on agent infrastructure, not a bigger model

The most telling signal from Code with Claude 2026 is what wasn’t shipped: no new frontier model, and instead a stack of agent plumbing — Managed Agents that run tools in sandboxes you control and reach private Model Context Protocol servers, plus MCP Tool Search landing inside Claude Code so agents discover tools on demand instead of front-loading every definition. The underlying advanced tool-use features (Tool Search, Programmatic Tool Calling) report an ~85% cut in token usage and large accuracy gains on big tool libraries (Opus 4.5: 79.5% → 88.1% on MCP evals). Why it matters for the lab: this is our architecture — agents over MCP services (BioEngine, Hypha, Agent-Lens). Context-frugal tool discovery is a direct, adoptable win.

🧫 The virtual cell gets a reality check

Scaling continues — Arc Institute’s STATE model, the Tahoe-100M perturbation atlas, and the community Virtual Cell Challenge — but 2025–26 benchmarking is refreshingly sober: multiple studies find deep perturbation-prediction models don’t yet clearly beat simple linear baselines, and a new 2026 preprint asks bluntly whether today’s AI virtual-cell models are actually useful for discovery. Why it matters for the lab: the virtual cell is squarely on our horizon — and the lesson is to pair ambition with rigorous baselines and honest, leakage-free benchmarks before trusting predictions.

🔬 Self-driving labs go mainstream

Nature ran a feature on the self-driving-lab revolution, and a wave of reviews (Materials Horizons’ “SDL 2.0”) and national-lab programs (Argonne, Sandia) show closed-loop AI + robotics moving from one-off demos to standing infrastructure — mostly in chemistry and materials, with biology close behind. Why it matters for the lab: REEF and our self-driving microscope live here; the design patterns — active learning in the loop, observe→reason→act, augmenting rather than replacing scientists — transfer directly.

🧬 Spatial omics gets its foundation models

The image × omics convergence we care about is crystallizing into foundation models: Nicheformer (single-cell and spatial, pretrained on a 110M-cell corpus), plus a fast-growing catalog of spatial/histopathology models (spEMO, scGPT-spatial, KRONOS) that read H&E images and molecular profiles together. Why it matters for the lab: models that jointly reason over microscopy images and omics are exactly the bridge between our bioimage-AI tooling and the cell biology it serves.

📖 From the lab

Fresh from us: Wei Ouyang and Hanzhao Zhang wrote the “Large Language Models and AI Agents” chapter for Janelia’s AI in Microscopy: A BioImaging Guide — a practical tour from deep learning to autonomous, agent-driven bioimage analysis. We wrote it up here.

Sources are linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. Spotted something we should cover? Nudge us.