The Rise of AI Agents in Biotech, Where Are We Now?

In this post, we take a grounded look at these emerging systems. No hype, just the current landscape: what’s working, what isn’t, and what technical opportunities and constraints are shaping the field today.

🤖 What Are AI Agents Doing in Biotech—Technically?

Modern agents in biotech combine three technical layers:

I. LLM Core: Typically GPT-4, Claude 3.5, or DeepSeek for task decomposition and natural language reasoning. Some use fine-tuned BioLLMs like BioGPT for domain context.

II. Tool Layer: Dynamically called APIs and function endpoints. These can include:

• Structured biomedical APIs (OpenFDA, ChEMBL, PubChem);
• Local database queries (e.g., from curated gene sets or lab-specific datasets);
• Code execution for data analysis using internal Python kernels.

III. Orchestration Framework: Chain-of-thought agents often use custom ReAct-style loops, where the LLM issues tool calls and updates an internal scratchpad. ToolRAG (used in TxAgent) performs retrieval-aware tool selection, for instance.

Agents like TxAgent and BioDiscoveryAgent are not monolithic—they rely on this orchestration model to dynamically pull in tools and retrievable biomedical context at each reasoning step.

🧭 Spotlight: SpatialAgent and the Frontier of Spatial Genomics

One of the most technically complete and domain-specific agent frameworks to emerge recently is SpatialAgent by a team of Genentech, developed to tackle spatial biology problems using a self-governing LLM architecture.

Spatial genomics—studying how genes are expressed across tissues in space—is computationally intensive and biologically complex. Existing methods are fragmented, require significant manual input, and often underperform on multimodal datasets.

SpatialAgent changes that by integrating LLM reasoning with domain-specific tools and structured task decomposition. It operates with three tightly integrated modules—memory, planning, and action: