Build a planner workflow

A high-tier planner agent creates a step-by-step plan; a pure-R dispatcher routes each step to the appropriate worker agent (no LLM call per dispatch); an optional evaluator agent decides whether the results are complete or require replanning.

Usage

planner_workflow(
  planner,
  workers,
  evaluator = NULL,
  max_replans = 2L,
  max_steps = 10L,
  parse_plan = NULL,
  state_schema = NULL
)

Arguments

planner

An Agent object used for planning (typically a high-capability model such as Opus).

workers

Named list of Agent objects that execute plan steps.

evaluator

An optional Agent object that reviews completed results. If NULL, the workflow ends when the plan is exhausted.

max_replans

Integer. Maximum number of replanning rounds (default 2L).

max_steps

Integer. Expected maximum plan length, used to size the termination guard (default 10L).

parse_plan

Optional function(text) that converts the planner's raw response into a list of list(worker, instruction) items. If NULL, a default line-by-line parser expecting worker_name: instruction is used.

state_schema

A WorkflowState or NULL (uses default).

Value

A compiled GraphRunner.

Details

Graph: START -> planner -> dispatcher -> [workers] -> dispatcher -> ... -> evaluator -> (done -> END | replan -> planner).

The planner must respond with one step per line in the format worker_name: instruction. A custom parse_plan function can be supplied to handle alternative formats.

Examples

if (FALSE) { # \dontrun{
runner <- planner_workflow(
  planner   = agent("planner",    ellmer::chat_anthropic(model = "claude-opus-4-6"),
                    instructions = "Break tasks into steps for 'researcher' and 'writer'."),
  workers   = list(
    researcher = agent("researcher", ellmer::chat_anthropic(model = "claude-haiku-4-5-20251001")),
    writer     = agent("writer",     ellmer::chat_anthropic(model = "claude-haiku-4-5-20251001"))
  ),
  evaluator = agent("evaluator",  ellmer::chat_anthropic(model = "claude-opus-4-6"))
)
result <- runner$invoke(list(messages = list("Write a short report on tidy data.")))
result$get("results")
} # }