What a New Agentic Chemistry Workflow Teaches Us About AI in Production

The news: an agent connected to real scientific tools

On November 3, 2026, Google DeepMind described a cheminformatics workflow built around Google Antigravity and a set of domain-specific skills. In its example, the system searched literature and compound resources for crizotinib- and lorlatinib-related compounds, curated the findings, compared matched molecular pairs, and explored potency models. The article says the workflow used RDKit alongside connectors to ChEMBL and PubChem, and presented results in an interactive viewer.

This is more interesting as an architecture pattern than as a claim that AI has replaced medicinal chemistry. The model coordinates steps and interprets results; domain software performs structured molecular operations; scientists assess whether the evidence is useful. DeepMind explicitly notes limitations from sparse data, assay noise, and narrow measurement ranges, and says proposed compounds still need experimental validation.

Pattern 1: let the model orchestrate, not redefine the domain

Chemical structures are not prose. A small change to a structure can change the compound itself. In this example, the agent used established cheminformatics tooling instead of asking a language model to calculate molecular properties from text. The same separation is valuable in finance, healthcare, and engineering:

• Use the model to translate intent into a bounded workflow.
• Use deterministic, domain-tested libraries for calculations and transformations.
• Validate inputs and outputs against domain schemas before they cross system boundaries.
• Keep provenance for every result: source, transformation, tool version, and assumptions.

The domain tool should remain the authority for operations where correctness is exact. The model can explain or prioritize results, but should not be the only validation layer.

Pattern 2: make source provenance part of the output

The demonstrated workflow extracts information from scientific literature and databases. A useful production system must let a reviewer move from a claim back to the source record. Preserve identifiers, source URLs, extraction confidence, relevant passages, and any normalization applied. Do not flatten these into an unattributed summary.

This makes review faster and helps distinguish a source error from a model error. It also makes refreshes safer: when an external record changes, the system can identify which derived results depend on it.

Pattern 3: separate exploration from consequential actions

Searching, filtering, and plotting are reversible exploration. Ordering a compound, changing a clinical record, or sending a customer communication is not. Give exploratory steps read-only tools and put irreversible actions behind explicit authorization and human review. Keep a clear boundary between a candidate hypothesis and an approved decision.

Pattern 4: evaluate the workflow, not just the final answer

A plausible narrative can hide a bad retrieval or an invalid intermediate calculation. Evaluate each stage: source retrieval, entity resolution, data cleaning, tool arguments, transformation correctness, model calibration, and the final handoff. Use known examples and holdout cases. Where historical data is available, prefer splits that reflect the real deployment question; a random split can make related examples appear in both training and evaluation and overstate generalization.

For scientific modeling specifically, predictive scores are not experimental evidence. Track the dataset version, split strategy, assay context, baseline, uncertainty, and failure cases with every model result.

A production checklist for domain agents

1. Define the narrow task and the authorized data sources.
2. Use deterministic libraries for domain calculations.
3. Keep tool schemas small, typed, and permission-scoped.
4. Attach source provenance to extracted facts and derived results.
5. Test malformed inputs, missing records, noisy data, and tool outages.
6. Require expert review before decisions with safety or financial impact.
7. Measure end-to-end task quality, not just latency or response fluency.

The strongest takeaway is not that every team needs a science agent. It is that useful agents are assembled around reliable domain capabilities, inspectable evidence, and a responsible human decision process.

Sources

• Google DeepMind: Agentic cheminformatics with Google Antigravity (November 3, 2026)
• RDKit: Open-source cheminformatics
• ChEMBL database
• PubChem