​
Overview

This guide covers common patterns and best practices for building effective Julep agents, inspired by a blog written by Anthropic on Building effective agents.

​
Core Workflow Patterns

​
1. Prompt Chaining

Break tasks into sequential steps, where each step’s output feeds the next. Can be used for:

  • Content generation and subsequent checks
  • Translation/localization in multiple stages
  • Ensuring quality between discrete transformations

Example implementation:

main:
# Step 1: Generate initial content
- prompt:
   role: system
   content: >-
      Generate marketing copy for product X based on the following:
      target_audience: {{_.audience}}
      product_features: {{_.features}}
      keywords: {{_.seo_keywords}}
   unwrap: true

# Step 2: Quality check gate
- evaluate:
   quality_check: _.content

# Step 3: Translation
- prompt:
   role: system
   content: Translate the approved content to Spanish
   unwrap: true

​
2. Routing Pattern

Act as a “traffic controller” by detecting request types and sending them to the correct handler. Ideal when inputs are diverse or require specialized expertise.

Steps generally include:

  1. Classification of incoming data
  2. Routing to handler modules
  3. Specialized processing
  4. (Optional) Aggregation of results

Example implementation:

main:
# Classification step
- prompt:
   role: system
   content: Classify the input query type
   unwrap: true

# Route based on classification
- switch:
   - case: _.classification == "technical_support"
      then:
         - tool: handle_technical
         arguments:
            ...
   - case: _.classification == "billing"
      then:
         - tool: handle_billing
         arguments:
            ...
   - case: _.classification == "general"
      then:
         - tool: handle_general
         arguments:
            ...

​
3. Parallelization Pattern

Execute subtasks concurrently by either dividing the workload (sectioning) or collecting multiple perspectives (voting).

Keys to consider: • Parallel processes for performance or redundancy • Syncing results and handling errors • Aggregating diverse outputs

Example implementations:

  1. Sectioning:
tools:
- name: aggregate_results
  type: ... # depends on the specific needs

# Custom workflow to run a subtask
run_subtask:
- ....

# Main workflow
main:
- prompt:
   role: system
   content: >
      Break this task into multiple subtasks.
      Here is the task: {{_.task}}
   unwrap: true

- over: _.subtasks
   do:
      - workflow: run_subtask
      arguments:
         ...
- tool: aggregate_results
  arguments:
    results: _
  1. Voting:
tools:
- name: perform_voting
  description: Perform voting on the results of running the task instances, and return the majority best result.
  type: ... # depends on the specific needs

# Custom workflow to run a subtask
run_subtask:
- ....

# Main workflow
main:
- over: _.main_tasks
   do:
      - workflow: run_subtask # Run the same task multiple times (given that the `run_subtask` workflow is non-deterministic)
      arguments:
         ...
- tool: perform_voting
  arguments:
    results: _

- evaluate:
   final_result: _

​
4. Orchestrator-Workers Pattern

Use a central “orchestrator” that delegates subtasks to multiple “worker” agents and integrates their outputs. Best for:

  • Large or dynamic multi-step tasks
  • Coordinating various specialized capabilities
  • Flexible task distribution

Example implementation:

main:
# Orchestrator planning
- prompt:
   role: system
   content: Break down the task into subtasks. Here is the task: {{_.task}}
   unwrap: true
# Worker delegation
- foreach:
   in: _.subtasks
   do:
      tool: assign_worker
      arguments:
         task: _

​
5. Evaluator-Optimizer Pattern

Create iterative feedback loops to refine outputs until they meet preset criteria. Suitable for: • Content refinement • Code reviews • Detailed document enhancements

General flow:

  1. Generate an initial result
  2. Evaluate against criteria
  3. Provide improvement feedback
  4. Optimize/retry until goals are met

Example implementation:

tools:
- name: score_content
  description: Score the content based on the criteria. Returns a json object with a score between 0 and 1, and a feedback string.
  type: function
  function:
    parameters:
      type: object
      properties:
        content:
          type: string
          description: Content to score

# Subworkflow to evaluate content
evaluate_content:
- tool: score_content
  arguments:
    content: _.content
- if: _.score < 0.5 # If the content does not meet the criteria, improve it
  then:
    - workflow: improve_content
      arguments:
        content: _0.content # _0 is the main input of this workflow
        feedback: _.feedback # _ is the output of the score_content tool call
  else:
    evaluate:
      final_content: _0.content

# Subworkflow to improve content
improve_content:
- prompt:
    role: system
    content: Improve the content based on the feedback. Here is the feedback: {{_.feedback}}
  unwrap: true
- workflow: evaluate_content
  arguments:
    content: _


main:
# Initial generation
- prompt:
   role: system
   content: Generate initial content. Here is the task: {{_.task}}
   unwrap: true

# Evaluation loop
- loop:
   while: "not _.meets_criteria"
   do:
      - tool: evaluate_content
      - tool: improve_content

Explanation:

  1. The evaluate_content subworkflow:

    • Takes content as input and scores it using a scoring tool
    • If the score is below 0.5, it triggers the improvement workflow
    • Uses special variables (_0 and _) to manage content and feedback between workflows
    • Returns the final content once quality criteria are met

  2. The improve_content subworkflow:

    • Receives content and feedback from the evaluation
    • Uses an LLM to improve the content based on specific feedback
    • Automatically triggers another evaluation cycle by calling evaluate_content

The main workflow ties these together by:

  • Generating initial content from a task description
  • Running a continuous loop that alternates between evaluation and improvement
  • Only completing when content meets the defined quality criteria

This creates a powerful feedback loop where content is repeatedly refined based on specific feedback until it reaches the desired quality level. The pattern is particularly useful for tasks requiring high accuracy or quality, such as content generation, code review, or document analysis.

​
Best Practices

Start Simple

  • 1. Minimal Steps: Use minimal steps and only add complexity when it clearly improves results.

Thoughtful Tool Design

  • 1. Documentation: Provide clear documentation, examples, and usage guidelines for each tool.

Error Handling

  • 1. Guardrails: Include feedback loops, define stopping conditions, and surface potential issues.

Testing

  • 1. Validation: Thoroughly test in controlled environments and measure against success criteria.

Human Oversight

  • 1. Checkpoints: Establish checkpoints for approval, ensure transparency, and maintain easy-to-audit workflows.

​
Conclusion

These patterns represent proven approaches from production implementations. Choose and adapt them based on your specific use case requirements and complexity needs.

​
Support

If you need help with further questions in Julep:

Architecture Deep DiveLocal Setup