📅 November 3, 2026 • deepmind.google
What happened
Google DeepMind has shipped Cheminformatics Skills for its Antigravity agent, announced on the DeepMind blog on November 3, 2026. The agent can now search across thousands of compounds, run matched molecular pair analysis and propose rigidified macrocycles with synthesis routes. The stated pace is 1 hour per iteration against the typical several days. DeepMind demonstrated the capability on real drugs targeting lung cancer.
How it works
The skills are built on RDKit, the open-source chemistry toolkit maintained for more than 20 years, with connectors to the ChEMBL and PubChem databases. The architecture splits the work explicitly. The LLM handles orchestration and reasoning, deciding which step comes next and interpreting what comes back. Deterministic chemistry tools handle the calculations themselves, from substructure search to molecular property computation. Every step the agent takes is a reproducible, checkable operation rather than a generated guess.
Why deterministic tools beat a pure LLM loop
An agent that computes with the same model that writes its plan has no ground truth to fall back on. Here the LLM never invents a molecular property. It calls RDKit for the number, and it can be wrong only in choosing the call, not in producing the value. That keeps the audit trail concrete. A chemist can replay the exact sequence of tool calls, inspect the inputs and verify each result against the toolkit itself. Errors localize to a step instead of hiding inside generated text.
IMPACT
The transferable pattern is agent plus domain skills. Any team automating a technical field can wrap its proven deterministic tooling as callable skills and let the model handle planning and interpretation. DeepMind reports the loop compresses chemistry iteration from days to about an hour. Whether that pace holds outside the demo set has not been independently benchmarked, but the architectural claim stands on its own. Orchestration by the LLM, arithmetic by verified tools.
Sources: deepmind.google, rdkit.org, ebi.ac.uk
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
