# 05 — The Agent Plumbing Layer: Biomni, OriGene, STELLA, LabOS, PantheonOS & Perturb-Seq Agents While systems like *The Virtual Biotech* and *AI Co-Scientist* focus on **macro-level reasoning architectures**, a parallel wave of 2025–2026 papers solved the **software engineering and wet-lab integration problem**: how do you give agents reliable access to hundreds of specialized bioinformatics packages, evolve new tools on the fly, select sequential Perturb-seq experiments under tight budgets, and physically supervise human hands at the wet bench? --- ## 1. Biomni: General-Purpose Biomedical Agent via Action Discovery (*Science*, 2026) * **Citation:** Huang, K., Zhang, S., Wang, H., Qu, Y., ..., Leskovec, J., et al. *"Autonomous biomedical research with an artificial intelligence agent."* ***Science*** (Published online August 20, 2026), DOI: `10.1126/science.adz4351`, PMID: `42424436` (Preprint: *bioRxiv*, mid-2025). * **Institutions:** Stanford University, Genentech, Arc Institute, UCSF, Princeton. ### Why It Matters for AI Engineers Bioinformatics doesn't have one clean SDK—it has thousands of brittle Python/R scripts, cryptic CLI tools, and 50+ fragmented web APIs. Manually writing MCP wrappers for every tool doesn't scale. ### Key Innovation: "Action Discovery Agent" + Code-as-Action 1. **Automated Environment Mining:** An *Action Discovery Agent* read thousands of recent *bioRxiv* papers across **25 biomedical domains** to catalog what tools and APIs real biologists actually use. 2. **Unified Action Space:** It compiled **`150` specialized biological tools**, **`105` software packages** (including single-cell staples like `scverse`, `Pertpy` for perturbation/CRISPR analysis, and `Squidpy` for spatial data), and **`59` databases** into a single executable environment. 3. **Code-as-Action + RL Variants (`Biomni-A / E / R`):** Rather than rigid JSON tool schemas, Biomni uses retrieval-augmented planning to write and execute Python/R pipelines dynamically—allowing zero-shot generalization across **causal gene prioritization** (benchmarked against Open Targets Genetics), **Perturb-seq / CRISPR screen design**, **rare disease diagnostics**, and **ADMET drug safety prediction**. --- ## 2. OriGene: Self-Evolving Virtual Disease Biologist + 600 MCP Tools * **Citation:** Zhang, Z., Qiu, Z., Wu, Y., Li, S., Wang, D., Zhou, Z., ..., Bai, L., Gao, Q., Sun, S., & Zheng, S. *"OriGene: A Self-Evolving Virtual Disease Biologist Automating Therapeutic Target Discovery."* ***bioRxiv*** (2025–2026). * **Institutions:** Shanghai Jiao Tong University, Fudan University, Lingang Laboratory, Westlake University, Shanghai AI Lab. ### Architecture * **600+ MCP Tools + Knowledge-Graph Tool RAG:** Integrates over 600 bioinformatics and pharmacology tools via the **Model Context Protocol (MCP)**. To avoid overwhelming the LLM context window with 600 tool schemas, OriGene uses a knowledge-graph-backed RAG router to retrieve only context-appropriate tools per step. * **Two-Level Self-Evolution:** 1. *Test-Time Iteration:* Coordinator, planning, reasoning, and critic agents recursively refine answers per query. 2. *Template Evolution:* Updates the system's reusable reasoning templates across generations based on empirical feedback. ### Validation (`TRQA` & Patient-Derived Organoids) * Introduced the **TRQA (Target Research Question-Answering)** benchmark covering target biology, clinical trials, and R&D strategy. * Nominated overlooked oncology targets—specifically **`GPR160` in liver cancer** and **`ARG2` in colorectal cancer**—which showed significant anti-tumor activity when subsequently tested in **patient-derived organoids (PDOs)** and tumor fragments. --- ## 3. STELLA & LabOS: Self-Evolving "Tool Oceans" + XR Wet-Lab Co-Pilots These two tightly linked Princeton/Stanford papers bridge **self-evolving software tooling** with **physical wet-lab perception**: ### 3A. STELLA (*arXiv:2507.02004*, July 2025) * **Authors:** Jin, R., Wang, M., et al. (Princeton / Stanford). * **Core Mechanism:** Features an evolving **Template Library** of reasoning patterns plus a dynamic **Tool Ocean**. When STELLA encounters a bio problem it lacks a tool for, a dedicated **Tool Creation Agent** autonomously searches, writes, tests, and registers new bioinformatics tools into the Tool Ocean. * **Results:** Doubles accuracy on complex tasks across iterative trial loops; hit **~26% on Humanity's Last Exam (Biomedicine)**, **~54% on LAB-Bench DBQA**, and **~63% on LAB-Bench LitQA** (up to +6 percentage points over static baselines). ### 3B. LabOS (*arXiv:2510.14861*, October 2025) * **Authors:** Cong, L., Zhang, Z., Wang, X., Di, Y., Jin, R., ..., Wu, J. C., & Wang, M. (Stanford School of Medicine & Princeton). * **The Engineering Leap:** LabOS connects STELLA's dry-lab multi-agent planner (`98+` dynamic bioinformatics tools) directly to the human benchtop via **Extended Reality (XR) smart glasses** and robotics. * **LabSuperVision (LSV):** Introduces a vision-language benchmark specifically for wet-lab perception—watching a scientist's hands through smart glasses in real time, catching pipetting/reagent errors before they ruin a multi-week stem-cell or cancer immunotherapy experiment, and logging exact physical state transitions back to the agent swarm. --- ## 4. PantheonOS: Evolving Bioinformatics Algorithms (*bioRxiv*, 2026) * **Citation:** Xu, W.-Z., Poussi, E., ..., Engreitz, J. M., & Qiu, X.-J. *"PantheonOS: An Evolvable Multi-Agent Framework for Automatic Genomics Discovery."* ***bioRxiv*** (2026). * **Institution:** Stanford University. * **Key Idea:** Uses a Four-Layer Pyramid Architecture paired with **`Pantheon-Evolve`**—an agentic code evolution loop that mutates and benchmarks genomics algorithms themselves (e.g., single-cell batch-correction and RL-augmented gene panel selection) against quantitative fitness metrics. Applied to uncover paracrine *Cer1–Nodal* inhibition in 3D spatial mouse embryo atlases and molecular heart-disease programs. --- ## 5. PerTurboAgent & SpatialAgent: Agents for Perturb-Seq & Spatial Biology Led by **Aviv Regev's group at Genentech** (with Stanford collaborators), these two agents tackle the two highest-dimensional experimental readouts in functional genomics: ### 5A. PerTurboAgent (Sequential Active-Learning Perturb-Seq) * **Citation:** Hao, M., Lee, Y., Wang, H., Scalia, G., & Regev, A. *"PerTurboAgent: A Self-Planning Agent for Boosting Sequential Perturb-seq Experiments."* ***bioRxiv*** & ***MLCB (PMLR)***. * **The Problem (Zero-Bio Context):** Single-cell Perturb-seq (editing genes + sequencing every single cell's transcriptome) is extremely expensive. You **cannot** test all 20,000 genes $\times$ combinations at once. Instead, labs run **sequential rounds** (test 200 genes $\rightarrow$ see what happens $\rightarrow$ pick the next 200 genes). * **How It Works:** PerTurboAgent combines an *action memory*, statistical analysis tools, and biological literature/pathway priors to decide which genes to CRISPR-perturb next. * **Results:** Outperformed standard mathematical active learning (`Coreset`, `AdversarialBIM`) on AUC hit curves—achieving a **hit ratio of `0.27` vs. `0.23`** on hard, sparse-hit phenotypes while excels on complex pathways (Translation, Proteasome, Mitochondrial Respiratory Chain). *(Note: the authors also highlight higher initial-run variance vs. deterministic heuristics, showing why feedback loops matter).* ### 5B. SpatialAgent (*bioRxiv*, April 2025) * **Citation:** Wang, H., Leskovec, J., Regev, A., et al. (Genentech / Stanford). * **What It Does:** Plan-Act-Conclude multimodal agent spanning custom gene-panel design, cell-type/niche annotation, spatial cell-cell communication (CCI), and trajectory inference—benchmarked across **2M+ cells** (human brain, fetal heart, tonsil, mouse colitis/prostate) and matching or beating human expert baselines. * **Verification Trend (`PROVER`):** Accompanying auditing work introduced claim-centric trace verification (`PROVER`) to audit whether cellular perturbation agents produce valid inferential links rather than plausible-sounding post-hoc stories. --- ## Key Engineering Takeaways * **MCP + Tool RAG is becoming standard (`OriGene`, `Biomni`):** Once you pass ~100 bio tools, dumping schemas into prompts fails; dynamic tool retrieval and `Code-as-Action` dominate. * **Self-evolving toolkits (`STELLA`, `Pantheon-Evolve`):** Static agent toolsets decay quickly in fast-moving genomics; the winning pattern lets a *Tool Creation Agent* write and unit-test new scripts on demand. * **Closing the active-learning loop (`PerTurboAgent`, `LabOS`):** Biology's verification cost means agents must act like **Bayesian experimental designers**—maximizing information gain per dollar of wet-lab sequencing or pipetting.