Chemistry / Synthesis Agents

This page describes the Chemistry / Synthesis Agents row of the Papers.md matrix: LLM agents specialized for chemistry, covering synthesis planning, molecular reasoning, and autonomous chemical experimentation. The row’s authoritative scope is its Taxonomy.md definition; this page synthesizes what currently sits in it.

Scope boundary

The row is for agents whose domain is chemistry, which is narrower than it first sounds. An agent that reads the chemistry literature belongs in Scientific Literature & Discovery Agents; one that drives instruments without a chemistry-specific tool layer belongs in Robot Scientists & Lab Automation; one that reasons over biological rather than chemical entities belongs in Domain-Specific Biomedical Agents. What puts a paper here is that the tools the LLM is given are chemistry tools: reaction prediction, retrosynthesis, molecular property calculation, safety screening.

Both references sit in the AI Tooling / Methodology column, because neither is applied to a cellular-agriculture problem. They are catalogued because the architecture transfers: media formulation, scaffold chemistry, and flavor-molecule design are all chemistry problems posed in a cell-ag setting, and these two papers are the reference implementations of an LLM planning and executing chemical work.

AI Tooling / Methodology

• #70 Coscientist (Boiko et al. 2023, Nature): a GPT-4 system that designs, plans and executes chemical experiments, demonstrated across six tasks including optimization of palladium-catalysed cross-couplings. The architecture is the durable contribution and is worth reading structurally: a Planner module holds the reasoning loop and is given a four-command action space, GOOGLE (web search, itself an LLM), PYTHON (calculation, run in an isolated Docker container so a generated script cannot reach the host), DOCUMENTATION (retrieval and summarization of an instrument API) and EXPERIMENT (execution against real hardware). The documentation and experiment commands were demonstrated against the Opentrons Python API and Emerald Cloud Lab’s Symbolic Lab Language, so the same pattern covers both a benchtop liquid handler and a cloud lab. The Planner can also repair its own code when execution errors come back, which is what turns a one-shot generation into a loop. Catalogued in Software.md.
• #71 ChemCrow (Bran et al. 2024, Nature Machine Intelligence): GPT-4 augmented with 18 expert-designed chemistry tools, organized as general, molecular, and reaction tools and wired together through LangChain on a ReAct-style reasoning loop. The agent planned and executed syntheses of an insect repellent and three organocatalysts, and guided discovery of a novel chromophore. Two details matter beyond the headline. The tool layer includes a literature-search tool built on paper-qa with OpenAI embeddings over a FAISS index, so the agent grounds answers in documents rather than in weights, which is the same retrieval pattern the literature-agent row is built around. And the pipeline includes a safety check that halts execution when an input molecule is identified as a controlled chemical, one of the few published agent designs where refusal is a component rather than a prompt instruction. Evaluation combined LLM-based and expert human assessment. Code at ur-whitelab/chemcrow-public.

Read together, the two papers separate a question that is easy to conflate: Coscientist shows an agent closing the loop to physical hardware, ChemCrow shows an agent whose competence comes from a curated tool set rather than from the base model. A cell-ag team building an autonomous formulation or assay loop is choosing along both axes at once.

Adjacent methods

• Robot Scientists & Lab Automation: the closed-loop experimental systems Coscientist’s EXPERIMENT command connects to, including work that predates the LLM era.
• Scientific Literature & Discovery Agents: the retrieval-grounded agents ChemCrow’s LitSearch tool is drawn from.
• General-Purpose Biomedical Agents and Domain-Specific Biomedical Agents: the biology-facing counterparts of the same tool-augmented pattern.
• Agent Infrastructure (Frameworks, KGs, Protocols): the framework and protocol layer these agents are assembled from.
• Reinforcement Learning: where chemistry reasoning models post-trained on verifiable problems sit, rather than in this row.

Further reading

• Research area: AI Tooling / Methodology, and the field overview in AIAgentsFoundationModels.md.
• Software: Software.md for Coscientist, and the AI Agents & Foundation Models section for the wider agent tooling.
• Evaluation: Benchmarks & Evaluation Frameworks, for how well agents of this kind actually perform.