Software agents find a pattern in old drug trials
Stanford Medicine reported on 17 September 2026 a virtual company of about 37,000 software agents that reviewed some 50,000 clinical trials. Where single-cell data existed, the agents scored how narrowly a drug hit a cell type and whether the target gene behaved more like a switch than a dimmer. Trials aimed at switch-like genes were, in that catalogue, 40 percent more likely to move from phase 1 to phase 2, 48 percent more likely to reach the market, and linked to 32 percent fewer adverse events than broadly active ones. Using only information from before January 2025, the agents sketched an antibody-drug design against B7-H3; a company later, in August 2025, independently arrived at a similar design that received an FDA breakthrough designation. The news report does not present the agents' own sketch as a medicine already tested in patients.
The record ends here. Everything below this line is invented.
What if a drug were chosen because its gene behaved like a switch?
Part one
The Switch Gene
Esme did not run a trial. She chose which genes were allowed to have one. The agents had read the old failures and come back with a blunt rule: prefer a target that is on or off in one kind of cell, not a dimmer spread through the tissue. Those were the drugs that had stumbled forward more often.
Two genes sat on her slate for the same tumor. One was everywhere, a dimmer, familiar, with a chemistry the plant already knew. The other was a switch in a fibroblast at the tumor's edge, a cell she had been taught to treat as scenery. The agents had drawn a homing tag for it and a payload, and they had drawn it from papers old enough that a real company had later walked the same path without them.
She was not allowed to call that a cure. She was allowed to spend the one slot the quarter would pay for.
Esme struck the familiar gene off the slate. The switch remained, untested, at the top.
What’s real
- About 37,000 agents catalogued some 50,000 trials
- Switch-like targets were linked to better trial odds in that catalogue
- A B7-H3 sketch later matched an independent company program
- The report does not show the agents' sketch tested in patients
What’s invented
- Esme's single paid slot for the quarter
- A familiar dimmer gene she strikes off
- A rule that the switch outranks the known chemistry
- A slate that treats the sketch as the candidate
Part two: The Gene She Kept
Part two isn’t on sale yet. Check back soon.
Filed under medicine, drug discovery, clinical trials, ai agents.