What Google showed on October 15
Google DeepMind described a Google Antigravity workflow for finding antimicrobial peptide candidates. In its account, a researcher steers an agent across protein records, sequence and structure comparisons, genomic context, and scientific literature. The system narrows a large search space and helps develop a hypothesis and a plan for experiments.
A useful agent still needs an evidence trail
The interesting engineering pattern is the chain of work: retrieve records, compare them, form a hypothesis, then decide what would test it. The agent’s answer should not flatten those steps into one confident paragraph. Keep the retrieved evidence, the model’s interpretation, and the next verification step distinct. For a production assistant, record which source and version supported a claim, what the model inferred, and whether a person or external process confirmed it.
More signals are not always independent proof
Different tools can produce useful, complementary clues, but a longer list is not automatically stronger evidence. Check where each result came from: two search results may repeat the same source, and a model summary may simply restate an earlier prediction. The scientific walkthrough reports several computational clues; the company still describes the candidate as awaiting laboratory validation.
Prediction and measurement answer different questions
DeepMind says the peptide has predicted activity and that it is undergoing wet-lab validation. That wording matters: database, sequence, structure, and literature analysis can help prioritize a candidate, while experiments test whether the predicted effect occurs under defined conditions. The source is a company walkthrough, not independent evidence that the candidate is safe, effective, or clinically useful.
Try the same evidence boundary in a support assistant
- Illustrative example: give a fictional support assistant a read-only copy of a return policy and ask whether a made-up order appears eligible. Keep all data fake; this is a software example, not medical guidance.
- For each response, keep the exact policy passage and its version alongside the assistant’s conclusion. Label a quoted rule as retrieved evidence and ‘this order seems eligible’ as an inference.
- Test a missing policy, two conflicting versions, an order with no purchase date, and a relevant policy passage that does not answer the question. The safe response may be to say what is unknown or request review.
- Before enabling an action, check that the evidence supports it and send uncertain cases to a human. Keep the prompt, retrieved passages, model output, and final action in the test record so failures can be reproduced.
The reported candidate is still being tested
DeepMind’s October 15 post says activity is predicted and validation is underway. It does not report a completed lab result, a safety assessment, or clinical usefulness. We did not use Antigravity, reproduce this workflow, or run an independent evaluation.