Skip to main content

Module 18: Parallel Processing with ParallelAgent

Theory​

ParallelAgent: Concurrent Execution​

The ParallelAgent is a workflow agent that executes all of its sub_agents at the same time (concurrently). This is a powerful pattern for improving performance when your workflow involves multiple independent tasks that can be performed simultaneously.

Key Concepts:

  • Execution Order: Not guaranteed. The agents run in parallel, and their execution may interleave. The ParallelAgent completes only when its slowest sub-agent has finished.
  • Shared State: All sub-agents in a ParallelAgent share the same session state. They can all read the initial state, and they can all write their results back to the state.
  • Avoiding Conflicts: When using a ParallelAgent, it's crucial to ensure that the sub-agents write their results to different keys in the state to avoid race conditions (where one agent's result overwrites another's).

When to Use ParallelAgent:

  • When tasks are independent and can run in any order.
  • When you need to gather data from multiple sources concurrently to save time.
  • When tasks do not need each other's outputs to run.

The Fan-Out/Gather Pattern​

A common and highly effective architecture is the fan-out/gather pattern, which combines the strengths of both parallel and sequential agents:

  1. Fan-Out (ParallelAgent): A ParallelAgent runs multiple data-gathering agents concurrently. Each agent is responsible for a specific, independent task (e.g., searching for flights, hotels, and activities).
  2. Gather (SequentialAgent): A final LlmAgent runs after the ParallelAgent has completed. Its job is to read all the results from the session state and synthesize them into a single, coherent response.

This gives you both speed (from parallel data gathering) and synthesis (from a final merging step).

        ┌──── Agent 1 (flights) ────┐
User ───┼──── Agent 2 (hotels) ─────┼──→ Merger Agent → Final Result
        └──── Agent 3 (activities) ─┘

      ParallelAgent (fast!)       SequentialAgent (combine)

Key Takeaways​

  • The ParallelAgent is a workflow agent that executes its sub-agents concurrently, improving performance for independent tasks.
  • Performance Note: The total execution time of a ParallelAgent is limited by the slowest of its sub-agents, including the latency of LLM calls and tool executions. This makes ParallelAgent particularly advantageous for workflows involving multiple, potentially slow LLM/Tool interactions, as they run simultaneously rather than sequentially.
  • It's crucial to use different output_keys for each sub-agent in a ParallelAgent to avoid overwriting data in the shared state.
  • The "fan-out/gather" pattern, which combines a ParallelAgent for data collection and a SequentialAgent for synthesis, is a powerful architecture for building efficient, complex agents.