Kratos: an auditable agentic workflow for automated calibration and population estimation of target-mediated drug disposition models

Abstract Building target-mediated drug disposition (TMDD) models involves many repetitive, error-prone steps. We developed Kratos, a proof-of-concept auditable agentic workflow in which a deterministic controller dispatches specialised agents for data validation, model specification, bounded multistart calibration with an adaptive rules planner, nlmixr2 FOCEi population estimation, diagnostics and reporting, and records every decision and numerical attempt. All analyses used the rules planner; no large language model (LLM) was called. Kratos was verified by cross-engine comparison, an automated test suite and recovery of known parameters from simulated data, and was applied to 143 published mean concentration–time profiles of 36 human monoclonal antibodies, split by molecule into two parts that were analysed independently and used to predict each other. On simulated data all 6 estimated typical parameters were recovered within 9% of truth, Python and R predictions agreed to a relative difference of 1.2 × 10 − 6 or less, and 54/54 automated tests passed. For the antibody data, linear disposition and absorption parameters were consistent across parts (clearance 0.188 and 0.268 L/day; bioavailability 0.71 and 0.84), whereas target-mediated parameters were not identifiable; the Part B covariance step failed and was flagged. Each part’s model predicted the other part’s unseen molecules with median fold errors of 1.29 and 1.35. The workflow ran from raw data to report without analyst intervention and kept numerical failures visible. Accuracy and reproducibility were supported; an efficiency advantage of adaptive over fixed multistart planning was not demonstrated.