Tool-call guardrail middleware for small and local models in multi-step agent workflows

Small and local models fail structurally in multi-step agentic workflows -- malformed JSON tool calls, wrong tool selected, plain text instead of structured output -- even when their reasoning is sound. Developers running cost-optimized pipelines (cheap worker models, expensive planner models) watch their agent loops derail repeatedly. Adding validation-and-retry guardrails around tool calls without retraining pushes reliability from 53% to 99% on multi-step tasks. No product fills this gap: existing solutions require model fine-tuning, backend-specific configuration, or only work with frontier models. The Forge open-source proof-of-concept showed the demand in May 2026 (687 HN points, 252 comments) but is a library, not a deployable product with monitoring, routing, and provider abstraction.