AutoFlow (Router workflow)

[jenna-winkler]

Description

This issue tracks the implementation of a workflow pattern where a Delegator step dynamically determines the next step to execute based on context or input intent. This pattern allows routing between different models/tools and can optionally return a final result directly to the user.

This approach enables:

• Dynamic control flow within workflows
• Modular task-specific steps using different models
• Returning to the Delegator for coordination after each step

Context

Originally discussed in Github Discussion #793

Next actions

• Implement the example in Python
• Update docs
• Review docs

(Nice to have)

• Create a TypeScript version if time permits

---

Implementation

Autoflow is a workflow engine where an LLM determines which functions to execute and in what order. Instead of having predetermined execution paths, the LLM acts as a router that decides the next step based on:

• The current state
• Available functions
• The user's request

How it works

1. Register functions as workflow steps
2. User provides a task
3. LLM examines available steps and current state
4. LLM decides which step to run next and what parameters to use
5. System executes the selected step
6. Process repeats until completion (when LLM calls the "final_answer" step)

Core components

• Autoflow Class is the main class for creating and executing workflows
• AutoflowState tracks the workflow's current execution state

Autoflow uses two main templates to guide the LLM:

• Router Template - used to help the LLM decide which step to run next
• Response Template - formats action results in the conversation

Example

https://github.com/i-am-bee/beeai-framework/tree/feat/803-autoflow-example/python/examples/workflows/custom/autoflow