Tool Chaining Pattern - Sequential Tool Execution

The Tool Chaining pattern enables LLMs to execute multiple tools in sequence, where each tool’s output becomes input for subsequent tools.

Overview

Pattern Description

Tool chaining allows complex workflows by:

1. Planning a sequence of tool executions
2. Passing outputs from one tool to the next
3. Building up results incrementally
4. Synthesizing final output from all results

Chain Phases

1. Chain Planning

The LLM analyzes the request and determines:

• Which tools are needed
• Execution order based on data dependencies
• How data flows between tools

2. Chain Execution

Tools execute sequentially:

• Each tool receives its required inputs
• Outputs are captured and passed forward
• Errors can trigger replanning

3. Result Synthesis

The LLM combines all outputs:

• Interprets results in context
• Generates natural language response
• Includes artifacts (files, charts, etc.)

Example Chain

User Request: “Analyze our Q4 sales and create a presentation”

Data Flow Patterns

Linear Chain

Tool A → Tool B → Tool C → Result

Fan-Out

        ┌→ Tool B ─┐
Tool A ─┤          ├→ Tool D
        └→ Tool C ─┘

Fan-In

Tool A ─┐
        ├→ Tool C → Result
Tool B ─┘

Conditional

Tool A → [condition] → Tool B (if true)
                    → Tool C (if false)

Chain Definition

{
  "chain_id": "sales_analysis",
  "steps": [
    {
      "id": "fetch",
      "tool": "query_database",
      "inputs": { "query": "SELECT * FROM sales WHERE quarter='Q4'" },
      "outputs": ["raw_data"]
    },
    {
      "id": "clean",
      "tool": "transform_data",
      "inputs": { "data": "${fetch.raw_data}" },
      "outputs": ["clean_data"]
    },
    {
      "id": "analyze",
      "tool": "calculate_metrics",
      "inputs": { "data": "${clean.clean_data}" },
      "outputs": ["metrics", "trends"]
    },
    {
      "id": "visualize",
      "tool": "generate_charts",
      "inputs": {
        "metrics": "${analyze.metrics}",
        "trends": "${analyze.trends}"
      },
      "outputs": ["charts"]
    }
  ]
}

Error Handling

When to Use

• Multi-step data processing workflows
• Tasks requiring incremental transformation
• Complex analysis requiring multiple tools
• Report generation pipelines
• ETL-style operations

Best Practices

Chain Design

• Keep chains as short as possible
• Make each step idempotent when possible
• Include validation between steps
• Cache intermediate results

Data Passing

• Use typed schemas for tool outputs
• Validate data between steps
• Handle missing/null values
• Transform data formats as needed

Performance

• Execute independent branches in parallel
• Stream large data when possible
• Set appropriate timeouts per step
• Monitor chain execution time

Benefits

Related Patterns