Show HN: Forge – Guardrails take an 8B model from 53% to 99% on agentic tasks | Hacker News

Forge, an open-source guardrail layer for self-hosted LLM tool use, adds retries, step enforcement, recovery, and memory-aware context handling.

In evaluations across many model and backend combinations, it dramatically improved multi-step reliability for local 8B models.

The big takeaway: much of the gap to frontier API agents may be closed by better reliability layers around the model, not just bigger models.

The results also show that serving backend and missing error-state handling can materially change tool-calling performance on consumer hardware.