SkillEffect: Checked Lowering for Memory-Bounded Agent Tools

Details

When agents call tools with large inputs, a semantically correct program can still exceed the memory allocated for a single tool call — crashing or degrading silently. SkillEffect addresses this with a checked-lowering runtime that requires every computation to define a source relation, a bounded implementation, and an output postcondition. Before execution authority is granted, an independent checker reconstructs each proposed lowering from the submitted program and immutable input. The architecture covers six operator families with five execution patterns, from streaming reduction to bounded-heap Top-k. Each relation plugin provides a recognizer, extractor, bounded-IR constructor, arena-bound function, and postcondition, while the dispatch, resource control, and publication mechanisms are shared. The XLSX onboarding study and Top-k extension show that new relations and retained-state patterns reuse the same trust boundary. Across evaluations, bounded access substantially reduces peak memory and improves completion rates under externally fixed memory caps. The checker accepts all legal configurations and rejects all adversarial proposals in the study. For platform engineers building agent tool infrastructure, SkillEffect offers a concrete contract model for memory-bounded execution — particularly relevant as agents increasingly handle large documents and datasets within single tool calls.