Design Philosophy of Workflow+limitedPlanning

Background

In practical AI Agent systems, many teams adopt a hybrid architecture: Workflow + Limited Planning, where an LLM performs lightweight planning followed by executing a predefined workflow.

Many real-world AI products use this architecture.
Most systems essentially implement Workflow + Limited Planning.

Examples:

LangGraph (state machine workflow)
Notion AI
Perplexity AI

These systems:

are not fully autonomous agents,
but rather:
structured workflows.

Workflow + Limited Planning Design Concept

User Input
      ↓
Intent Detection
      ↓
Task Planner
      ↓
Workflow Selection
      ↓
Workflow Execution
      ↓
Result

Detailed Process Flow (Complete Example)

Assume the user input:

Help me write an AI industry report

System process:

Step 1 Intent Detection

The LLM first determines the user’s intent:

intent = research_report

For example, output:

{
 "intent": "research_report"
}

Step 2 Planner (Lightweight Planning)

Planner decides:

which workflow is needed

For example:

research_workflow

Step 3 Workflow Selection

The system selects a predefined workflow:

Research Workflow

Structure:

Search Data
↓
Extract Key Info
↓
Summarize
↓
Write Report

Step 4 Workflow Execution

Each step calls the LLM or tools:

Search → Web API
Extract → LLM
Summarize → LLM
Write → LLM

Common Implementation Patterns of Limited Planning

Different systems implement this differently, but typically there are 3 patterns.

Pattern 1 Router Agent (Most Common)

The planner is only responsible for:

selecting which Agent to engage

Structure:

User Input
      ↓
Router
  ├ Research Agent
  ├ Coding Agent
  └ Writing Agent

Example:

User input:

Write a Python scraper

Router:

coding_agent

Pattern 2 Tool Selection

Planner only chooses tools:

Search
Database
Code Interpreter

Example:

User: Search AI companies’ funding

Planner:

use web_search

Pattern 3 Workflow Routing

Planner selects a workflow:

content_workflow
analysis_workflow
coding_workflow

Then executes the fixed flow.

Implementation at Code Level (Simplified)

A typical Python structure:

def router(user_input):

    intent = llm_intent_detection(user_input)

    if intent == "research":
        return research_workflow(user_input)

    elif intent == "coding":
        return coding_workflow(user_input)

    elif intent == "summary":
        return summary_workflow(user_input)

Workflow:

def research_workflow(query):

    data = search(query)

    info = extract(data)

    summary = summarize(info)

    report = write_report(summary)

    return report

Here:

planning = router

not a full agent.

References

ChatGPT

• Previous
  AutoGPT Analysis and Autonomous Thinking
• Next
  AutoGPT分析と自律的思考