RcrewAI Rails

Examples

End-to-end recipes for the most common patterns.

Batch execution

Run a crew once per input set; each input becomes its own Execution, grouped by a shared batch_id.

crew = RcrewAI::Rails::Crew.find_by(name: "research_team")

result = crew.execute_batch_sync([
  { topic: "AI in healthcare" },
  { topic: "AI in finance" }
])

result[:batch_id]                        # shared id for the batch
crew.batch_executions(result[:batch_id]) # the two Executions, in order

A guardrailed task

Validate and transform a task's output with a host class.

class JsonGuardrail
  # Must return [ok, value_or_error]
  def validate(output)
    parsed = JSON.parse(output)
    [true, parsed]
  rescue JSON::ParserError => e
    [false, "Invalid JSON: #{e.message}"]
  end
end

task = crew.tasks.create!(
  description: "Return the result as JSON",
  expected_output: "A JSON object",
  guardrail_class: "JsonGuardrail",
  guardrail_method_name: "validate",
  guardrail_max_retries: 3
)

Knowledge-backed agent

Give an agent reference material to draw on.

agent = crew.agents.create!(name: "analyst", role: "Analyst", goal: "Answer from the docs")
agent.knowledge_sources.create!(source_type: "pdf", value: Rails.root.join("docs/handbook.pdf").to_s)
agent.knowledge_sources.create!(source_type: "url", value: "https://example.com/reference")
# Sources are embedded lazily when the crew runs.

A persisted Flow

Flow subclasses live in your app; the engine persists state and tracks runs.

class OnboardingFlow < RCrewAI::Flow
  start :greet
  def greet
    state.step = "greeted"
  end
end

run = RcrewAI::Rails::FlowRun.execute(OnboardingFlow, inputs: { user_id: 42 })
run.status  # "completed"
run.result  # { "step" => "greeted", ... }

A consensual crew

Have multiple agents propose and vote on each task's answer.

crew = RcrewAI::Rails::Crew.create!(
  name: "review_board",
  process_type: "consensual",
  consensus_agents: 3
)
# add agents + tasks as usual, then:
crew.execute_sync