From seven combination hypotheses to one testable interaction: a gated agentic AI QSP workflow applied to healthy ageing interventions

Large language model (LLM) agents can propose combination therapies and construct supporting mechanistic models far quicker than either can be verified. To address this gap, we built a gated agentic AI quantitative systems pharmacology (Ai QSP) workflow where no hypothesis reaches a report until it clears strict hurdles for precedent, evidence, structural integrity, and release. We applied this pipeline end to end to healthy-ageing interventions. It began with a 34 state model calibrated on clinical trial data for semaglutide and metformin. Locked and evaluated against an independent post hoc DNA methylation trial, the model captured the direction of change across all 9 organ and system clocks, but the magnitude was far too small: it predicted about 0.14 years per year where 1.6 were observed, halting tier promotion. The workflow then generated seven nutritional combination hypotheses (H1H7) with 48 new parameters encoded in SBML. Identifiability analysis showed that 37 of these were completely unconstrained by available data. Trimming the model down to its lead pairing - a food derived bioactive (Agent A) plus a single-strain probiotic (Agent B left six identifiable parameters and locked the uncalibrated interaction term at zero. From this reduced model, we derived an 84 day trial with a 28 day washout, requiring 356 participants for 80% power at an interaction ratio of 0.75. A 9,000 person virtual trial confirmed the design (80.5% power), but revealed an operational trap: an assay floor at 24 mg/kg slashed power to 33% and skewed estimates toward the null, forcing a protocol-level censoring condition before study lock. Testing the pipeline across downstream operational stages and comparative oncology programmes showed that unconstrained generation consistently over-reports novelty, topological entry points can masquerade as biological mechanism, and governance gates remain vulnerable to operator overrides. Ultimately, the workflow delivered no silver bullet: its output is a single testable interaction, an assay hardened trial design, and an audit trail documenting why the other six hypotheses were shelved. Agent identities, doses, and agent specific sources are withheld in this version pending patent filing.