LLMRagent: agents, multi-agent conversations, and agent experiments

LLMRagent builds on LLMR's provider layer to provide:

Details

• agent(): an agent with a persona and an LLMR model config. It calls native tools, keeps memory, holds to a budget, and can stream replies with chat(stream = TRUE).

• agent_as_tool(): expose an agent as a tool, so other agents can delegate to it; this is the primitive behind supervisor/specialist hierarchies.

• agent_pipeline(): run text through a fixed chain of specialists, keeping every intermediate product.

• conversation(): multi-agent conversations over a shared, attributed transcript, with turn-taking policies and stop rules; presets debate(), focus_group(), interview(), and deliberate().

• agent_experiment(): factorial designs over conditions and replications, run sequentially or in parallel, returning one tidy results frame.

• think_harder(): an orchestrator that uses one strong model to plan and synthesize while many cheap models draft answers in parallel.

Every run yields a tidy transcript and a trace of calls, tokens, and timings. Combine with LLMR::llm_log_enable() for a complete per-call audit file.

See also

Useful links:

• https://github.com/asanaei/LLMRagent

• https://asanaei.github.io/LLMRagent/

• Report bugs at https://github.com/asanaei/LLMRagent/issues

Author

Maintainer: Ali Sanaei sanaei@uchicago.edu