Calibrating Trust in Small Local LLM Agents: Design Patterns for Research Workflows

A practice-based paper on how an LLM agent should signal its state so that a researcher's trust follows the agent's actual reliability. It uses small, quantized, open-weight models running on the researcher's own machine as a test bench, and proposes three principles (competence-match, transparency, control) and fourteen design patterns, each defined by the signal the agent must emit.