Forge – Guardrails take an 8B model from 53% to 99% on agentic tasks

Forge adds a guardrails wrapper around self-hosted LLM tool-calling without touching the underlying model. The headline number — 53% to 99% task completion on an 8B model — is the kind of claim that makes you pull up the repo at midnight. The mechanism is practical rather than magical: retry nudges when a step stalls, enforced step sequencing so the model cannot skip ahead and break the chain, error recovery hooks, and VRAM-aware context management so you are not silently truncating context on a 24GB card. The domain-agnostic design means it is not tied to a single task type, which matters for builders running heterogeneous agent workflows. Honest reservation: the benchmark is the author's own, not a third-party eval, so the 99% figure deserves some skepticism until you run it on your own task distribution. But even half that improvement on a self-hosted 8B model is worth a Saturday afternoon. -> Best for: AI engineer running self-hosted LLM agents on constrained hardware