LLMRagent

Reproducible Language-Model Agents for Research

An 'R' interface built on 'LLMR' creates large language model (LLM) agents for use as reproducible and governed research instruments. Agents pair a model configuration with optional persona instructions. Stateful exchanges retain conversational memory, and native 'R' functions can serve as tools within declared budgets. Several agents can hold a turn-taking conversation over a shared transcript, and factorial experiments across such designs run in parallel. Model calls and tool activity are captured in a run object, so the research record extends beyond generated text. A study manifest hashes the design and computational apparatus without treating sampled replies as part of its identity, and a hash-sealed archive preserves it alongside transcripts and call records. Tool policies record declared side effects and can limit call counts or result sizes; calls marked for human review pause before execution. Robustness checks assess sensitivity to prompt or model changes, while calibration against human labels corrects estimates drawn from imperfect model labels.

Versions across snapshots

Dependencies (latest)