Agent Infrastructure (Frameworks, KGs, Protocols)

This page describes the Agent Infrastructure (Frameworks, KGs, Protocols) row of the Papers.md matrix: the substrate agents run on, covering agent frameworks, biomedical knowledge graphs, tool and communication protocols such as MCP, and orchestration ecosystems. The row’s authoritative scope is its Taxonomy.md definition; this page synthesizes what currently sits in it.

Scope boundary

The discriminator is that the contribution is plumbing rather than an agent that solves a downstream task. A paper releasing a tool registry, a graph server, a training environment or a protocol belongs here even when it demonstrates an agent built on top; a paper whose point is what the agent achieved belongs in one of the four agent rows. #93 CellForge is the boundary case: it designs neural architectures for perturbation prediction, which is a downstream task, but its stated core contribution is the multi-agent framework itself, so it sits in both this row and Domain-Specific Biomedical Agents.

For a cell-ag team this is the least glamorous and most load-bearing row. None of these papers does biology. All of them determine whether an agent that does biology can be built without reimplementing tool calling, entity resolution and knowledge retrieval from scratch.

AI Tooling / Methodology

Tool ecosystems and protocols

#41 ToolUniverse (Gao et al. 2025, arXiv, Zitnik lab): standardizes how an AI scientist finds and calls tools, across more than 600 machine-learning models, datasets, APIs and scientific packages. The core is an AI-tool interaction protocol in three parts: a specification schema describing each tool’s function, parameters and outputs so any client can invoke it without knowing the implementation; an interaction schema encoding calls uniformly; and communication methods, running local tools directly in Python and remote ones over the Model Context Protocol. That abstraction makes a local function, a remote model and a laboratory instrument interchangeable. Six components cover the lifecycle: a Tool Finder using keyword, LLM in-context and embedding search; a Tool Caller that validates inputs and loads tools on demand; a Tool Manager for registration; a Tool Composer chaining tools sequentially, in parallel or with feedback; a Tool Discoverer that generates whole new tools from natural-language descriptions and validates them; and a Tool Optimizer that improves tool specifications by generating test cases and analysing executions. Demonstrated on a hypercholesterolemia case study that identified a potent drug analog. Code at mims-harvard/ToolUniverse; catalogued in Software.md.
#67 (Ruscone et al. 2025, bioRxiv, Saez-Rodriguez lab): builds MCP servers over three modelling tools, NeKo for constructing gene-regulatory networks from prior-knowledge databases, MaBoSS for Boolean-model simulation and PhysiCell for multicellular agent-based models, more than 60 specialized tools covering data collection through simulation setup. A multiscale model of cancer cell fate under TNF was built entirely through natural-language interaction. The three design principles the authors derive are the transferable output and apply to any lab wrapping its own tools: tool granularity should be defined at biological decision points where domain knowledge guides the modelling choice rather than at API boundaries; session management is essential for tracking long workflows and reproducing them; and orchestration has to be established from the start so the LLM can combine tools flexibly without losing biological coherence. Running the same task across three LLMs produced consistent end-to-end orchestration but model-dependent differences in network size and dynamical behaviour, which the authors present as evidence of both portability and variability.
#133 BioContextAI (Kuehl et al. 2025, Nature Biotechnology): a community registry of biomedical MCP servers plus reference implementations, so agentic assistants can discover and compose tool servers rather than each project integrating each resource itself. Its Knowledgebase MCP exposes more than 15 biomedical resources and is installable from PyPI or reachable as a hosted endpoint, with a cookiecutter template for new servers and published evaluation scripts. The paper’s framing puts humans in a supervisory role across an iterative agent-human loop covering hypothesis generation, experimental design and analysis. Catalogued in Software.md.

Knowledge graphs and graph serving

#48 BioChatter / BioCypher (Lobentanzer et al. 2025, Nature Biotechnology, Saez-Rodriguez lab): an open-source Python framework for building biomedical LLM software, whose explicit purpose is eliminating the repeated reimplementation of the same basic workflows. It harmonizes the APIs of open-source deployment tools and proprietary providers so a user can switch model or vendor without changing code, integrates knowledge graphs and vector databases for retrieval-augmented generation, and makes live web APIs usable by letting the LLM parameterize queries against per-API implementations. Deployments span a deliberate tradeoff axis from simple and economical to fully self-hosted and secure, with two interfaces (a Python “Light” app for prototyping and a fuller JavaScript one). Code at biocypher/biochatter; data at Zenodo.
#162 SciAtlas (Qiao et al. 2026, arXiv): a multidisciplinary academic knowledge graph built from OpenAlex, covering 26 disciplines, 43.30 million papers, 109.70 million authors, 3.76 million keywords and 0.12 million institutions, with 157 million entities and roughly 3 billion triplets across 11 relationship types. Papers are organized on four levels: semantic (citation and relatedness edges), conceptual (keyword co-occurrence), directional (domain, field, subfield, topic hierarchies) and social (co-authorship and institutional affiliation). Construction normalizes and deduplicates entity names but deliberately does not deduplicate authors, because name ambiguity makes it unsafe. The retrieval algorithm is neuro-symbolic, with tri-path collaborative recall and graph reranking, aimed at replacing keyword or vector matching with deterministic association discovery, and the paper argues this cuts the inference cost and logical hallucination of agentic deep-research frameworks. Code at zjunlp/SciAtlas.
#219 ESCARGOT (Matsumoto et al. 2025, Bioinformatics, Moore lab): couples an LLM to biomedical knowledge graphs through a dynamic Graph of Thoughts, targeting the two failure modes of plain RAG, context-length limits and imprecise vector similarity. The workflow is worth following because it is unusually explicit: generate three candidate strategies incorporating the graph’s real node and edge types, score them for clarity, comprehensiveness and schema alignment, convert the winner to Python, compile that into an XML graph where nodes are code-execution tasks and edges the logical flow, then traverse and execute, entering a self-debugging loop on any node failure and falling back to strategy regeneration if the XML will not compile. Knowledge is pulled via Cypher queries or RAG and turned into Python objects operated on directly, which keeps the LLM out of the extraction path. The authors emphasize that both the generated code and the knowledge requests can be vetted, unlike a black-box RAG answer. Code at EpistasisLab/ESCARGOT.
#220 PloverDB (Glen et al. 2025, Bioinformatics): serves Biolink-compliant knowledge graphs as TRAPI-compliant web APIs, and is included because the unglamorous problem it solves recurs in every graph-backed agent. It is fully in-memory Python with no Neo4j or SQLite underneath, its core structure a nested adjacency dictionary mapping nodes to neighbours organized by neighbour category and edge type, shared read-only across uWSGI workers. It handles entity resolution, standardized metadata and test-data exposure, and multiplexing of several graphs in one platform, configured by a single JSON file and deployed as a Docker image that downloads and indexes the graph at build time. Originally built in 2021 to replace Neo4j behind RTX-KG2 for query-speed reasons, generalized in 2024 to any Biolink graph. MIT licensed; code at RTXteam/PloverDB. It is the serving layer under the same RTX-KG2 graph #223 reasons over.

Training environments and orchestration

#160 Aviary (Narayanan et al. 2024, arXiv, FutureHouse): an extensible gymnasium for language agents, formalizing an agent as a policy solving a language-grounded partially observable Markov decision process, which the authors name a language decision process. Five environments ship, three of them scientific: manipulating DNA constructs for molecular cloning, answering research questions against the literature, and engineering protein stability. The headline result matters for anyone costing an agent deployment: with online training and scaled inference-time compute, agents backed by open-source non-frontier LLMs matched or exceeded both frontier-LLM agents and human experts on several tasks at up to 100 times lower inference cost. Apache-2.0; code at Future-House/aviary; catalogued in Software.md.
#61 Agentic Lab (Wang et al. 2025, bioRxiv): described under Robot Scientists & Lab Automation. It is in this row for its orchestration layer, a LangGraph state machine with a supervisor scheme, agent-as-tool interfaces between the virtual PI and its subagents, structured state management for continuity across multi-step workflows, and explicit safeguards against infinite loops and context overflow.
#93 CellForge (Tang et al. 2026, arXiv, Gerstein lab): autonomously designs neural architectures for perturbation prediction from raw multi-omics data and a task description, going beyond hyperparameter tuning to produce genuinely new components such as trajectory-aware encoders and perturbation diffusion modules. Its Task Analysis module parses experimental metadata across scRNA-seq, scATAC-seq and CITE-seq, then retrieves literature by combining a fixed set of 46 single-cell perturbation articles with live PubMed search alternating breadth-first layers (retrieving diverse concepts) and depth-first layers (following promising paths), scored by Sentence-BERT cosine similarity. Its Design module runs a graph-based expert discussion in which specialized agents (Dataset Analyst, Problem Investigator, Baseline Assessor) iteratively refine proposals and update confidence scores based on each other’s evaluations. Evaluated on six datasets spanning gene knockouts, drug treatments and cytokine stimulations, with reasoning traces retained throughout. Code at gersteinlab/CellForge; catalogued in Software.md.

Adjacent methods

General-Purpose Biomedical Agents and Domain-Specific Biomedical Agents: the agents this substrate carries.
Scientific Literature & Discovery Agents: built on the same retrieval and orchestration layers, with SciAtlas and ESCARGOT aimed squarely at their hallucination problem.
Robot Scientists & Lab Automation: where the protocol layer reaches physical instruments.
GNN: learned message passing over graphs, as opposed to the curated and queried graphs here.