Principle:CrewAIInc CrewAI Crew Execution
| Knowledge Sources | |
|---|---|
| Domains | Orchestration, Multi_Agent_Systems |
| Last Updated | 2026-02-11 00:00 GMT |
Overview
The execution phase that triggers a composed crew to process its tasks in order, interpolating runtime inputs, managing agent execution loops, and collecting results.
Description
Crew Execution is the runtime phase where the assembled crew begins processing. Input variables are interpolated into task descriptions, agents execute tasks in the configured process order (sequential or hierarchical), tool calls are managed, memory is updated, and results are aggregated. The execution engine handles error recovery, context window management, telemetry emission, and lifecycle callbacks.
Usage
Use this principle after assembling a crew with agents and tasks. Execution is triggered with optional input dictionaries that fill template variables in task descriptions and expected outputs. The result is a structured output containing raw text, optional structured data, and per-task outputs.
Theoretical Basis
Crew Execution implements a Pipeline Execution Model where tasks are processed sequentially with output from earlier tasks available as context for later ones. Each agent operates in a ReAct loop (Reason-Act-Observe) until the task is complete. The execution engine manages the lifecycle events, error boundaries, and state transitions.
Pseudo-code:
# Abstract execution flow
interpolate_inputs(tasks, input_dict)
for task in tasks:
context = gather_context(previous_task_outputs)
result = task.agent.execute(task.description, context, tools)
store_output(result)
return aggregate_results(task_outputs)