LASR-MCP: Learned Adaptive Schema Resolution with Dynamic In-Tree Recovery for Enterprise Tool-Calling Agents

Enterprise tool-calling agents face a schema-resolution trade-off: exposing a complete JSON Schema contract preserves formal constraint information but increases context overhead, while aggressive pruning can omit domain constraints required for valid execution. We present Learned Adaptive Schema Resolution (LASR-MCP), a three-tier mechanism combining (1) an empirically selected invariant foundation, (2) proactive task-conditioned constraint prediction, and (3) validator-driven pass-level multi-error in-tree JSON Schema recomposition. We evaluate LASR-MCP on a synthetic enterprise-style supply-chain benchmark comprising 32 evaluation tasks across four operational archetypes. On an unseen template family, the frozen semantic predictor achieves F1 = 0.863 versus 0.250 for a frequency-prior baseline (ΔF1 = +0.612, p < 0.001). In the 32-task development/validation benchmark, reactive recovery increases final structural success from 78.1% to 100.0% and reduces total provider-reported workload tokens by 20.2% at k = 1 and 22.1% at k = 2 relative to full-schema control. A separate post-freeze robustness study across four heterogeneous enterprise-style schemas reports 20.5–25.2% token savings and 16/16 ambiguous-task structural recovery. The results provide empirical evidence for the feasibility of constraint-level adaptive schema resolution within an already selected tool schema. Broader transfer across model families, APIs, and production deployments remains future work.