Skip to content

How do agentic AI and foundation models meet?

Explainer page for the FM-to-Virtual-Cells talk. Answers the question someone asks during Act 1 or Act 4: "what's the relationship between LLM agents and biology FMs?" The short answer: it's bidirectional. The longer answer matters for project selection because Lane 9 — the most commercially fundable lane in the 2026 menu — sits exactly at this intersection.

The headline

Agentic AI and biology FMs are complementary, not competing. There are four load-bearing patterns at the intersection:

  1. Agents that use FMs as tools — LLM agent calls a biology FM (frozen) the way it would call a calculator. PathChat-DX, BioAgents, MedAgentGym.
  2. Agents that build FMs — LLM agent + AI coding agent autonomously designs and trains a virtual-cell architecture. VCHarness (BioMap, 2026).
  3. Agents that reason over FMs — LLM post-trained with reinforcement learning where a biology FM serves as the verifier of biological plausibility. rBio (CZ Biohub, 2025).
  4. Agents that analyze data with FMs as substrate — autonomous comp-bio agent that runs a full analysis and generates new biological insight. CellVoyager (Nat Methods 2026).

All four exist as public 2025–2026 systems. The intersection is no longer vapor — and the Li et al. 2026 Nat Biotech position paper "Agentic AI and the rise of in silico team science" is the first venue-level framing of the whole intersection as a team-science shift, not just a tooling shift.

Interactive — the four patterns don't form an orthogonal 2×2; the public 2024–2026 systems run along one diagonal, from FM-consumers (bottom-left) to FM-producers (top-right). Hover any bubble for the canonical exemplar.

Pattern 1: Agents using FMs as tools

The structure: an LLM (Claude / GPT / open Llama / Qwen) sits at the top of the stack. It receives a user query in natural language ("Stage this tumor"; "Predict expression after CRISPRi on Gene X"). It decides which downstream tool to call, calls a biology FM (pathology / single-cell / genomic) as a remote function, then reasons over the FM's output to produce a final answer.

Canonical exemplars:

  • PathChat / PathChat-DX (Mahmood Lab, Nature 2024) — multimodal conversational pathology model. The LLM-side handles the natural-language interface; pathology-FM-side (UNI / CONCH) handles the slide encoding. PathChat-DX is the first generative-AI pathology tool with FDA Breakthrough Designation (Jan 2025) — the regulatory template for this pattern.
  • MedAgentGym (ICLR 2026, AACR-relevant) — 72k-task sandboxed code-exec gym with Med-Copilot-7B orchestrating biomedical analyses. The agent writes code to call scikit-learn, run scanpy, query databases. The substrate for training agents that use bio-FMs as tools.
  • Owkin Pathology Explorer in Anthropic's Claude HCLS environment (2026) — agent that uses Owkin's pathology FMs inside a Claude-based clinical workflow. The agentic interface for FM-driven pathology analysis.

Why this pattern matters: it makes biology FMs accessible to clinicians and biologists who can't write PyTorch. The interface layer is the LLM; the heavy lift is the biology FM. The agent doesn't fix the FM's accuracy problems — but it makes them tractable in a clinical workflow.

Pattern 2: Agents building FMs

The structure: an LLM agent + AI coding agent (think Claude Code or Cursor on biology) autonomously designs and trains a virtual-cell architecture from scratch. The agent reads the dataset, proposes an architecture, writes training code, runs the training loop, evaluates the result, and iterates. The output is a trained model, not a predicted answer.

Canonical exemplar:

  • VCHarness (Cheng et al., bioRxiv 2026.04.11) — BioMap Research + MBZUAI (Le Song + Eric Xing). Combines an AI coding agent with multimodal biological FMs to construct perturbation-response models. Identifies architectures that outperform expert designs; reduces development time months → days.

Why this pattern matters: it changes who can build a virtual cell. A lab without ML engineering capacity can describe what they want and let the agent prototype. The pattern is early — VCHarness is one paper from one lab as of April 2026 — but it's the agentic ↔ virtual-cell builder direction made concrete. Pairs with the rBio reasoning direction (Pattern 3): rBio reasons over an existing virtual cell; VCHarness builds new ones.

Pattern 3: Agents reasoning over FMs

The structure: an LLM is post-trained with reinforcement learning. The reward signal comes from a biology FM serving as a verifier of biological plausibility — not from RLHF on human preferences, not from text-based correctness, but from "does the biology FM agree this is true?"

Canonical exemplar:

  • rBio v1 (CZ Biohub, 2025) — Qwen2.5-3B-Instruct post-trained via GRPO (Group Relative Policy Optimization) using TranscriptFormer as a soft-supervision verifier. First reasoning model trained on virtual-cell simulations. Answers questions like "Would suppressing gene A increase activity of gene B?" with biological grounding from a sc-FM rather than from web text.

Why this pattern matters: it inverts the standard LLM tool-use direction. Instead of the LLM asks the FM for an answer, here the LLM is shaped by the FM during training. The output is an LLM whose reasoning reflects the FM's encoded biology. This is the virtual-cell reasoning model the field had been asking for since the 2024 Bunne Cell perspective.

Caveat: rBio is post-trained on Qwen2.5-3B — a small LLM. Performance is limited by both the LLM's capacity and the verifier's accuracy. If the verifier (TranscriptFormer) is wrong on a query, rBio learns to be confidently wrong. This is the same risk profile as RLHF-with-bad-rewards, transposed to biology.

Pattern 4: Agents analyzing data with FMs as substrate

The structure: an autonomous comp-bio agent receives a dataset and a loose goal ("find what's interesting here"). It runs a full analysis pipeline — QC, clustering, differential expression, pathway enrichment — calling FMs and classical tools as needed, and surfaces new biological insight rather than a single predicted value. This is the agent as a junior computational biologist, not as a calculator.

Canonical exemplar:

  • CellVoyager (Alber et al., Nat Methods 2026) — "AI CompBio agent generates new insights by autonomously analyzing biological data." Where VCHarness builds a model and rBio reasons over one, CellVoyager does the analysis — autonomously, end-to-end, on real single-cell data.

Why this pattern matters: it's the pattern closest to a working scientist's daily loop. The other three patterns produce models or answers; CellVoyager produces analyses — the actual unit of comp-bio work. For a small lab, this is the pattern that most directly threatens-or-augments the bottleneck (analyst bandwidth). The Li et al. 2026 Nat Biotech "in silico team science" framing is essentially "what happens to a lab when Pattern 4 is reliable."

Caveat: an autonomous analysis agent that surfaces "insights" inherits the multiple-comparisons problem at machine speed. CellVoyager-style outputs need the same statistical hygiene a human analyst's would — and arguably more, because the agent can generate a hundred plausible-looking findings before a human would notice it's p-hacking.

The 2026 commercial frame

The intersection is where pharma + AI-native biotechs spend money:

  • JPM 2026 Day 2: AstraZeneca acquires Modella AI — multimodal pathology FMs + agentic AI capabilities in one stack. The agent-uses-FM-as-tool pattern at acquisition scale.
  • JPM 2026 Day 1: Lilly + NVIDIA $1B AI Co-Innovation Lab — infrastructure commitment that pre-supposes agentic interfaces on FMs.
  • 2026: Owkin launches Pathology Explorer inside Claude — agent ↔ FM as the commercial interface.

The AACR 2026 AT02 session "Agentic AI as the Cancer Researcher" + the 4/22 "Agentic AI as the Oncologist" session are the field's first conference-level commitment to this intersection on a clinical-AI stage.

What this means for academic project selection

Lane 9 (FM-aided experimental design / active learning) is the lane that sits at this intersection. It uses the FM-as-tool pattern (Pattern 1) in a closed loop with wet-lab experiments. The buyer is AI-native biotech (Recursion, Insitro, Latent Labs, Vevo) — they actively pay for this work.

Concrete project shape: partner with a wet-lab running CRISPRi or drug-perturb screens. Implement an LLM-orchestrated FM-guided selection loop (Pattern 1 — FM-as-tool). Compare to literature-prior baseline. Output: methods paper in Nat Methods / Cell Systems + a clinical-relevance paper.

The next-step research questions that the four patterns expose:

  • Pattern 1 (FM-as-tool): how does an agent decide which biology FM to call for a given query? Tool-routing for biology is unsolved.
  • Pattern 2 (FM-builder): VCHarness designs architectures but not pretraining objectives. Can an agent design a better objective than next-gene-prediction? See Track 2 in the supplementary.
  • Pattern 3 (FM-as-verifier): rBio uses one verifier (TranscriptFormer). Can a reasoning agent use multiple verifiers (TranscriptFormer + pathology FM + AlphaGenome) for cross-modality biological reasoning? This is the open follow-up — and the first paper to do it owns the citation.
  • Pattern 4 (FM-as-analysis-substrate): CellVoyager produces autonomous analyses — but who audits them? A statistical-hygiene layer for agentic comp-bio outputs (multiple-comparisons control, automated sensitivity analysis) does not exist. That layer is itself a publishable small-lab project.

Common misconceptions

"Agentic AI is replacing FMs." False. Agents need FMs as substrates. Without the biology FM, the agent has nothing to call.

"Agentic AI fixes the linear-baseline reckoning." No. The reckoning is about the FMs themselves; agents make those FMs accessible but don't improve their underlying accuracy. If scGPT loses to a linear baseline on perturbation prediction, an agent calling scGPT will also lose to that baseline.

"Agents need GPT-4-class capacity." Not necessarily. rBio is built on Qwen2.5-3B (small open model). MedAgentGym's Med-Copilot is 7B params. The bottleneck for biology agents is the verifier accuracy and tool quality, not the LLM size.

"This is too early to be a project area." False as of May 2026. VCHarness, rBio, MedAgentGym, PathChat-DX, Owkin Pathology Explorer are all 2024–2026 public systems. The agentic ↔ FM intersection has tools, papers, and commercial signals. It just doesn't have a settled taxonomy yet — which is the small-lab opportunity.

Where to go next

Last updated 2026-05-13.