Local tool calling hype meets developer reality

The "agentic" capabilities of 20B-35B local models remain fragile in practice, despite community hype and strong benchmark scores.

• –Small models often lack the reasoning depth to course-correct when a tool call fails, leading to infinite execution loops
• –Hallucinated success—where a model claims it created a file without actually calling the tool—is a common failure mode when context windows lose terminal outputs
• –Real-world local agent workflows still require heavy guardrails, strict JSON parsing, and specialized fine-tunes rather than relying on out-of-the-box chat interfaces
• –The user frustration points to a broader tooling gap, showing that platforms like Open WebUI and LM Studio need better native validation for autonomous actions