Closing the loop: agents, scheduling and optimization run the lab
Goal: cancer drug pairs that work better together. Thousands of dose pairs, days per test: the lab tests a few, learns, picks the next. No human in the loop.
DeckFastest layout within the agent's labware rules.
CameraDeck set up as designed?
Dry runSimulate the plan 6 times.
The lab loop
solid: the loop · dashed: side paths
Optimizerpicks the next platesBayesian opt.Schedulertimes every robot stepCP-SATRobotsarm, liquid handler, incubatorsimulatedCamerachecks each plate movedepth + labelsReadercounts surviving cells384 wellsModellearns synergy over dosesGaussian processRecoveryfailure → replanrulesAgentbad data → new limit, drop bad platesClaude
What you see
dark = better together
Synergy: the pair beats each drug alone. It hides in a small patch of the map.
Stepswhat happens nowWorkcellthe arm moving platesMachinesliquid handler, reader, incubator, agentsSchedulewhich machine does what, whenDiscoverybest guess vs. the true best pairResults, Datascores vs. other runs, the run's records