ChemCrow

Implements a ReAct-style agent that decomposes chemistry queries into subtasks and routes them to domain-specific tools (molecular property prediction, reaction planning, literature search). Uses LangChain's agent executor with chemistry-domain tools as a tool registry, enabling multi-step reasoning where the LLM decides which chemistry tools to invoke based on intermediate results and task requirements.

Wraps RDKit and other chemistry libraries as callable tools within the agent framework, enabling the LLM to request molecular property calculations (logP, molecular weight, TPSA, etc.) without direct code execution. The agent parses SMILES strings or chemical names, invokes the wrapped tools, and receives structured property outputs that feed into downstream reasoning.

Integrates chemistry-specific reaction planning tools (e.g., retrosynthesis engines, reaction databases) into the agent framework, enabling the LLM to decompose target molecule synthesis into reaction sequences. The agent queries reaction databases, evaluates synthetic feasibility, and generates step-by-step synthesis routes with intermediate molecules and required reagents.

Integrates chemistry literature search and knowledge retrieval tools (e.g., PubChem, ChemSpider, arXiv chemistry papers) into the agent framework, allowing the LLM to query scientific literature for reaction conditions, property data, and synthesis precedents. The agent retrieves relevant papers or database entries and extracts structured information to inform chemistry decisions.

Manages state across multi-step chemistry workflows where outputs from one tool become inputs to the next (e.g., generate molecule → predict properties → check synthesis feasibility → retrieve literature). Uses LangChain's memory and state management to track intermediate results, maintain context across agent steps, and enable backtracking or alternative paths when tools fail.

Provides a conversational interface where chemists can describe chemistry tasks in natural language, and the agent translates these descriptions into tool calls and structured chemistry operations. The LLM acts as a semantic parser, converting phrases like 'find the most drug-like molecule' into sequences of property calculations and filtering operations.

Includes chemistry-domain prompts and few-shot examples that guide the LLM's reasoning about chemistry tasks, improving tool selection accuracy and reducing hallucinations. The agent uses chemistry-specific system prompts that establish domain context, define tool semantics, and provide examples of correct chemistry reasoning patterns.

Implements validation layers that check chemistry tool outputs for chemical validity (e.g., valid SMILES, chemically feasible reactions, reasonable property values) and gracefully handle tool failures. When tools return invalid results, the agent can retry with different parameters, fall back to alternative tools, or request clarification from the user.

Provides a plugin architecture where new chemistry tools can be registered and integrated into the agent without modifying core code. Tools are defined with schemas that describe inputs, outputs, and semantics, allowing the agent to discover and invoke tools dynamically. This enables integration with external chemistry software (molecular dynamics, quantum chemistry packages, etc.).

Manages conversation history and chemistry-specific context (molecules, reactions, properties) across agent interactions, enabling the agent to reference previous results and maintain coherent multi-turn chemistry conversations. Uses LangChain's memory abstractions with chemistry-specific serialization for molecules and reactions.