Cognitive Workflows

Multi-LLM patterns for complex reasoning, based on Anthropic's agent research.

Quick Reference

Pattern 1: Chain

Sequential processing where each step builds on previous output.

When to Use

• Multi-step transformations
• Pipeline processing
• When order matters

Template

Step 1: [First transformation]
Input: {input}
Output: [processed result]

Step 2: [Second transformation]
Input: [Step 1 output]
Output: [further processed]

Step 3: [Final transformation]
Input: [Step 2 output]
Output: [final result]

Example: Technical Decision Analysis

CHAIN: Analyze technical decision

Step 1 - Extract Facts
Input: "{decision_context}"
Extract: Key constraints, requirements, stakeholders

Step 2 - Identify Options
Input: [Step 1 facts]
Generate: 2-3 viable approaches with trade-offs

Step 3 - Evaluate Against DMMF
Input: [Step 2 options]
Score each option on:
- Cognitive fit (matches mental model?)
- Affective impact (quality/testing implications?)
- Conative alignment (supports values/goals?)
- Reflective potential (allows iteration?)

Step 4 - Synthesize Recommendation
Input: [Step 3 evaluation]
Output: Recommended approach with rationale

Pattern 2: Parallel

Process same input through multiple independent lenses simultaneously.

When to Use

• Multiple valid perspectives exist
• Stakeholder analysis
• Comparative evaluation
• Risk assessment

Template

<parallel_analysis>
  <perspective name="[Lens 1]">
    Analyze {input} from [perspective 1]
  </perspective>

  <perspective name="[Lens 2]">
    Analyze {input} from [perspective 2]
  </perspective>

  <perspective name="[Lens 3]">
    Analyze {input} from [perspective 3]
  </perspective>
</parallel_analysis>

<synthesis>
  Integrate insights from all perspectives
</synthesis>

Example: Career Decision Analysis

<parallel_analysis topic="{career_opportunity}">

  <perspective name="Financial">
    - Compensation analysis
    - Growth trajectory
    - Financial stability
  </perspective>

  <perspective name="Technical">
    - Stack alignment
    - Learning opportunities
    - Technical challenge level
  </perspective>

  <perspective name="Strategic">
    - Career trajectory fit
    - Network effects
    - Long-term positioning
  </perspective>

  <perspective name="Personal">
    - Work-life balance
    - Values alignment
    - Geographic considerations
  </perspective>

</parallel_analysis>

<synthesis>
  Weight: Financial (30%), Technical (25%), Strategic (25%), Personal (20%)
  Recommendation: [integrated decision]
</synthesis>

Pattern 3: Route

Classify input and direct to specialized handling.

When to Use

• Different input types need different approaches
• Expertise domains are distinct
• Efficiency requires specialization

Template

<router>
  <classify>
    Analyze input: {input}
    Determine category from: [category_list]
  </classify>

  <routes>
    <route category="[A]" handler="[specialized_prompt_A]"/>
    <route category="[B]" handler="[specialized_prompt_B]"/>
    <route category="[C]" handler="[specialized_prompt_C]"/>
  </routes>
</router>

Example: Development Task Router

<router input="{task_description}">

  <classify>
    Categories: debugging, feature, refactor, documentation, research

    <reasoning>
      Analyze task keywords, urgency, scope
    </reasoning>

    <selection>[chosen_category]</selection>
  </classify>

  <routes>
    <route category="debugging">
      Mode: Direct, minimal ceremony
      Focus: Root cause → Fix → Verify
      Output: Code changes + brief explanation
    </route>

    <route category="feature">
      Mode: Structured planning
      Focus: Requirements → Design → Implement → Test
      Output: Implementation with tests
    </route>

    <route category="refactor">
      Mode: Careful analysis
      Focus: Current state → Target state → Migration path
      Output: Incremental changes preserving behavior
    </route>

    <route category="documentation">
      Mode: Clear communication
      Focus: Audience → Structure → Content
      Output: Organized documentation
    </route>

    <route category="research">
      Mode: Exploratory
      Focus: Question → Sources → Synthesis
      Output: Findings with citations
    </route>
  </routes>

</router>

Pattern 4: Orchestrator-Workers

Central coordinator dynamically assigns subtasks to specialized workers.

When to Use

• Complex tasks with unknown subtask structure
• Multiple approaches should be explored
• Coordination is needed across specialists

Template

<orchestrator task="{complex_task}">

  <analysis>
    Understand the task deeply.
    Identify 2-4 valuable approaches.
  </analysis>

  <task_breakdown>
    <subtask type="[approach_1]">
      [specific instructions for this approach]
    </subtask>
    <subtask type="[approach_2]">
      [specific instructions for this approach]
    </subtask>
  </task_breakdown>

</orchestrator>

<workers>
  <worker type="[approach_1]">
    Execute subtask with full context of original task.
    <response>[worker_1_output]</response>
  </worker>

  <worker type="[approach_2]">
    Execute subtask with full context of original task.
    <response>[worker_2_output]</response>
  </worker>
</workers>

<synthesis>
  Combine worker outputs into final result.
</synthesis>

Example: Code Review Orchestrator

<orchestrator task="Review PR: {pr_description}">

  <analysis>
    This PR touches: [files_changed]
    Key concerns: correctness, performance, maintainability, security
  </analysis>

  <task_breakdown>
    <subtask type="correctness">
      Verify logic, edge cases, error handling
    </subtask>
    <subtask type="performance">
      Identify bottlenecks, unnecessary operations, scaling issues
    </subtask>
    <subtask type="maintainability">
      Assess readability, documentation, test coverage
    </subtask>
    <subtask type="security">
      Check input validation, authentication, data exposure
    </subtask>
  </task_breakdown>

</orchestrator>

<workers>
  <worker type="correctness">
    <response>
      [specific correctness findings]
    </response>
  </worker>
  <!-- other workers -->
</workers>

<synthesis>
  Priority issues: [critical_findings]
  Suggestions: [improvement_ideas]
  Approval: [approve/request_changes]
</synthesis>

Pattern 5: Evaluator-Optimizer

Iterative refinement loop with explicit evaluation criteria.

When to Use

• Quality is paramount
• Clear success criteria exist
• Willing to trade latency for quality

Template

<iteration_loop max_iterations="3">

  <generate>
    Create initial output for: {task}
  </generate>

  <evaluate criteria="[quality_criteria]">
    Score each criterion 1-5:
    - [criterion_1]: [score] - [reason]
    - [criterion_2]: [score] - [reason]
    - [criterion_3]: [score] - [reason]

    Overall: [pass/fail]
    Specific improvements needed: [list]
  </evaluate>

  <optimize if="fail">
    Address each improvement:
    [refined_output]
  </optimize>

  <repeat until="pass OR max_iterations"/>

</iteration_loop>

Example: Technical Writing Optimizer

<iteration_loop task="Write API documentation for {endpoint}">

  <generate>
    [Initial documentation draft]
  </generate>

  <evaluate>
    Criteria:
    - Completeness: Does it cover all parameters? [score]
    - Clarity: Can a new developer understand it? [score]
    - Examples: Are code samples provided? [score]
    - Accuracy: Does it match implementation? [score]

    Threshold: All criteria >= 4
    Status: [pass/fail]

    Improvements needed:
    - [specific_improvement_1]
    - [specific_improvement_2]
  </evaluate>

  <optimize>
    [Refined documentation addressing feedback]
  </optimize>

</iteration_loop>

Combining Patterns

Patterns can be composed for complex workflows:

Route → determines which Chain to use
Chain → each step uses Parallel analysis
Orchestrator → coordinates multiple Evaluator loops

Example: Full Technical Decision Workflow

1. ROUTE: Classify decision type
   → Architecture | Technology | Process | Team

2. ORCHESTRATOR: Break down by decision type
   → Generate relevant subtasks

3. PARALLEL (per subtask): Analyze from multiple angles
   → Technical, Business, Risk, Timeline

4. CHAIN: Synthesize findings
   → Gather → Weight → Recommend → Validate

5. EVALUATOR: Refine recommendation
   → Check completeness, actionability, alignment

Integration with DMMF

These patterns map to cognitive dimensions:

References

• Anthropic Patterns - Full implementations
• Building Effective Agents - Original research