Skip to contents

Changelog

Source: NEWS.md

LLMRagent 0.8.0

A reproducibility, governance, validity, and scale overhaul. The core agent is unchanged; every run now funnels through one unified object, and new layers wrap that object with provenance, control, defensible inference, and an auditable workflow runtime. Depends on LLMR (>= 0.8.7).

Provenance and reproducibility

  • [as_agent_run()] converts any run (an Agent, [conversation()], [debate()], [focus_group()], [interview()], [deliberate()], [agent_pipeline()], [agent_experiment()], or [think_harder()]) into one unified agent_run: a rich event graph of utterances, model calls, tool calls, state transitions, and budget/guardrail events.
  • tibble::as_tibble(run, level = ) exposes five tidy views of a run: "utterance", "event", "call", "tool", and "state".
  • [agent_manifest()] builds a study manifest of content hashes (persona, tools, workflow) plus the environment, via [hash_persona()], [hash_tool_spec()], and [hash_workflow()]; the hashes interoperate with LLMR.
  • [archive_agent_study()] writes a sealed, replayable archive whose call log is replayable with [replay_run()].
  • diagnostics() and report() (LLMR generics) gain agent_run methods for integrity checks and human-readable summaries.

Governance and control

  • [agent_tool()] declares a tool’s side effects (side_effects =), approval needs (requires_approval =), and quotas (max_calls =, timeout_s =, max_bytes =) up front; the manifest hashes them, so a writing tool cannot pass as inert.
  • [guardrail()] and [guardrails()] attach input, output, and tool checks to an agent (agent(guardrails = )); a rejection raises llmragent_guardrail_block.
  • [human_gate()] wraps a tool so a call pauses for a person; the pause raises llmragent_pending_approval carrying a checkpoint that [approve_tool_call()] and [resume_run()] act on.
  • [check_state_leakage()] diagnoses cross-condition contamination in an experiment.

Validity

  • [persona_frame()] builds a provenanced persona (source =, scope =, attributes =); [persona_variants()] generates counterfactual sets and [persona_audit()] scans for essentialism and stereotype.
  • [mark_claim_type()] enforces claim-type discipline and [llm_claim_lint()] flags claims unsupported by a run’s evidence; an unsupported assertion raises llmragent_claim_error.
  • [agent_calibrate()] bridges imperfect labels to a defensible estimate via design-based and prediction-powered estimation (method =, estimand =); [attach_calibration()] binds it to a run, and a "calibrated_inference" claim requires it. [as_llmrcontent_validation()] coerces to an LLMR content validation.
  • [agent_robustness()] runs a robustness battery (vary =, measure =) with [vary_models()], [vary_temperature()], [vary_prompt()], [vary_persona()], and [vary_option_order()].

Workflows and scale

  • [agent_workflow()], [add_node()], [add_edge()], and [run_workflow()] add a small, auditable DAG runtime; every node boundary is a checkpoint, so [resume_workflow()], [fork_workflow()], and [replay_run()] reproduce or branch a run, and a divergent replay raises llmragent_replay_mismatch. [workflow_from_pipeline()] lifts an [agent_pipeline()] into a workflow.
  • [mcp_tools()] is a governed Model Context Protocol client (policy =, transport =); it defaults to read-only with schema pinning and raises llmragent_mcp_schema_drift when a server’s schema changes.
  • [sandbox_tool()] runs a tool under isolation (mode =, executor =); a breach raises llmragent_sandbox_violation.
  • [agent_population()], [society()], [step_interaction()], [collect_measures()], [exposure_matrix()], and [contamination_report()] scaffold social simulation; [view_run()] renders an HTML run inspector.

Breaking changes

  • [interview()] now returns a classed agent_interview that carries provenance; use as.data.frame() for the old tibble.
  • [agent_experiment()] now returns a classed agent_experiment rather than a bare results frame.
  • [save_agent()] and the .rds format changed: spans replace the former trace field, so archives written by 0.7.x do not round-trip.

LLMRagent 0.7.1

A ground-up rewrite on LLMR (>= 0.8.7). The package now centers on one R6 Agent and a small set of composable layers above it.

Agents

  • agent(): persona + LLMR::llm_config() + tools + memory + budgets. chat() is stateful, reply() stateless, ask_structured() returns schema-shaped answers parsed into R lists.
  • chat(stream = TRUE) prints the reply token by token as it is generated.
  • Budgets (budget()) are checked before every call; the call that would exceed a limit raises a typed llmragent_budget_error instead of spending. Budget integrity is end to end: tool loops report every internal model call and its tokens (via LLMR’s tool_loop attribute), max_tool_calls is enforced inside a running tool loop (not only between turns), memory compaction calls land on the agent’s meter, and [load_agent()] restores the counters so a budget keeps binding across sessions.
  • Failures are errors, never replies: an API error propagates and leaves memory untouched.
  • trace() records every call, tool run, memory compaction, and budget stop with tokens and timings; usage() totals them.
  • ask_structured() no longer sends the agent’s tools alongside a schema request, avoiding provider conflicts between tool-choice and schema mode.

Delegation and pipelines

  • agent_as_tool() exposes an agent as an LLMR::llm_tool(), so other agents can consult it: supervisor/specialist hierarchies with attributed spend and nested budgets.
  • agent_pipeline() chains agents into an assembly line; every intermediate product is kept in a tidy steps frame.

Memory

  • Three drop-in policies: memory_buffer() (last n), memory_summary() (auto-compacts older history into a summary note; optionally billed to a dedicated cheaper model via config =), and memory_recall() (embedding-based retrieval of relevant older exchanges).
  • save_agent() / load_agent() round-trip an agent (config keys stay environment references; tools re-attach at load).

Conversations and study presets

  • conversation(): shared, speaker-attributed transcripts; round-robin, random, or moderator-chosen turn order; stop rules.
  • Presets returning tidy, classed objects with print and as.data.frame() methods: debate() (phased transcript + structured verdict), focus_group() (rotating speaking order + moderator synthesis), interview() (scripted questions + adaptive probes), deliberate() (discussion + private structured votes + tally).

Experiments and orchestration

  • agent_experiment(): factorial designs (conditions x replications), sequential or parallel via future, per-cell error capture, one tidy results frame.
  • think_harder(): one strong model plans, synthesizes, and verifies while many cheap models work the approaches in parallel; all intermediate products are retained.