Show HN: Forge – Guardrails take an 8B model from 53% to 99% on agentic tasks (github.com)

Forge has been introduced as a reliability layer for self-hosted large language models (LLMs), enhancing an 8 billion parameter local model's ability to execute multi-step, agentic tasks from a modest 53% success rate to an impressive 99%. This advancement is made possible through features such as rescue parsing, retry nudges, step enforcement, and meticulous context management, which collectively facilitate improved task execution in complex scenarios. The leading configuration, using Ministral-3 8B Instruct Q8 on llama-server, scored 86.5% across 26 different evaluation scenarios, showcasing its significant effectiveness. This announcement is particularly noteworthy for the AI/ML community as it addresses the ongoing challenge of reliability in LLMs during tool-calling workflows. Forge provides three key usage options: WorkflowRunner for defining and managing structured tasks, a middleware guardrails system that validates and corrects tool interactions, and a drop-in proxy server that seamlessly integrates with existing OpenAI-compatible clients. This flexibility allows developers to harness the full potential of LLMs while maintaining control over their operational logic, making it a critical tool for anyone looking to build robust AI applications that require dependable model performance in real-world tasks.