Workflows & Task Decomposition

How MUXI breaks complex requests into multi-agent workflows

Ask ChatGPT Ask Claude.ai Open in Cursor

Complex tasks need multiple agents with different skills. MUXI automatically decomposes requests into workflows, coordinates execution, and synthesizes results - no manual orchestration needed.

The Problem

Traditional approach:

User:  "Research AI trends, write blog post, create social media posts"

You manually:
1. Call researcher agent
2. Wait for results
3. Pass to writer agent
4. Wait for blog post
5. Pass to social media agent
6. Combine everything

Manual orchestration, error-prone, slow.

MUXI approach:

User:  "Research AI trends, write blog post, create social media posts"

MUXI automatically:
1. Analyzes complexity
2. Creates workflow
3. Executes in parallel
4. Returns combined results

Automatic orchestration, reliable, fast.

How It Works

graph TD
    U[User Request] --> O[Overlord]
    O --> C{Complexity Check}
    C -->|Simple ﹤7| A1[Single Agent]
    C -->|Complex ≥7| W[Create Workflow]
    W --> D[Decompose Tasks]
    D --> T1[Task 1]
    D --> T2[Task 2]
    D --> T3[Task 3]
    T1 --> E[Execute]
    T2 --> E
    T3 --> E
    E --> S[Synthesize Results]
    A1 --> R[Response]
    S --> R

graph TD
    U[User Request] --> O[Overlord]
    O --> C{Complexity Check}
    C -->|Simple ﹤7| A1[Single Agent]
    C -->|Complex ≥7| W[Create Workflow]
    W --> D[Decompose Tasks]
    D --> T1[Task 1]
    D --> T2[Task 2]
    D --> T3[Task 3]
    T1 --> E[Execute]
    T2 --> E
    T3 --> E
    E --> S[Synthesize Results]
    A1 --> R[Response]
    S --> R

Complexity Scoring

The Overlord scores requests from 0-10:

Score	Example	Action
0-3	"What's the weather?"	Simple, single agent
4-6	"Explain quantum computing"	Moderate, may use tools
7-10	"Research, analyze, create report"	Complex, trigger workflow

Configure threshold:

overlord:
  workflow:
    auto_decomposition: true
    complexity_threshold: 7.0  # Workflows trigger at 7+

Start with a higher threshold (8+) and lower it if needed. Too many workflows slow things down. Only decompose when the task genuinely needs multiple agents.

Workflow Creation

When complexity ≥ threshold:

User:  "Analyze our codebase, create security audit, file GitHub issues"
         ↓
Overlord analyzes:
  - Multiple steps mentioned
  - Different skills needed (analysis, security, GitHub)
  - Multiple output artifacts
         ↓
Complexity score: 9/10
         ↓
Create workflow:
  Task 1: Analyze codebase → code-reviewer agent
  Task 2: Security audit → security-agent (depends on Task 1)
  Task 3: Create issues → github-agent (depends on Task 2)

Workflow Components

1. Task Decomposition

Break request into subtasks:

Original: "Research competitors and create comparison report"

Decomposed:
  - Task 1: Research company A (researcher)
  - Task 2: Research company B (researcher)
  - Task 3: Research company C (researcher)
  - Task 4: Compare features (analyst)
  - Task 5: Create report (writer)

2. Dependency Management

Determine task order:

Task 1, 2, 3: Can run in parallel (no dependencies)
Task 4: Depends on 1, 2, 3 (needs research results)
Task 5: Depends on 4 (needs comparison)

3. Agent Assignment

Match agents to tasks:

Tasks 1-3: researcher (has web-search tool)
Task 4: analyst (has data-analysis tool)
Task 5: writer (has document-creation tool)

4. Parallel Execution

Run independent tasks simultaneously:

Start: Tasks 1, 2, 3 execute in parallel
         ↓
    All complete
         ↓
Start: Task 4 executes (uses results from 1, 2, 3)
         ↓
    Complete
         ↓
Start: Task 5 executes (uses result from 4)
         ↓
    Complete
         ↓
Return combined result to user

Workflow Example

Simple Request (Score: 4)

User:  "What's the capital of France?"

Overlord:
  - Score: 4 (factual question)
  - Route to: assistant agent
  - No workflow needed

Response: "The capital of France is Paris."

Complex Request (Score: 9)

User:  "Research top 5 project management tools, compare features,
       create a recommendation report"

Overlord:
  - Score: 9 (multiple steps, research, analysis, creation)
  - Create workflow:

    Task 1: Research Asana
      Agent: researcher
      Tools: web-search
      Output: feature_list_asana

    Task 2: Research Monday
      Agent: researcher
      Tools: web-search
      Output: feature_list_monday

    Task 3: Research ClickUp
      Agent: researcher
      Tools: web-search
      Output: feature_list_clickup

    Task 4: Research Jira
      Agent: researcher
      Tools: web-search
      Output: feature_list_jira

    Task 5: Research Trello
      Agent: researcher
      Tools: web-search
      Output: feature_list_trello

    [Tasks 1-5 run in parallel]

    Task 6: Compare features
      Agent: analyst
      Tools: data-analysis
      Input: feature_lists from Tasks 1-5
      Output: comparison_matrix

    Task 7: Create recommendation report
      Agent: writer
      Tools: document-creation
      Input: comparison_matrix from Task 6
      Output: recommendation_report.pdf

Response: "Here's your report comparing the top 5 PM tools..."
         [Attached: recommendation_report.pdf]

Execution Flow

sequenceDiagram
    participant U as User
    participant O as Overlord
    participant R as Researcher
    participant A as Analyst
    participant W as Writer

    U->>O: Complex request
    O->>O: Score complexity: 9
    O->>O: Create workflow
    O->>R: Tasks 1-5 (parallel)
    R-->>O: All research complete
    O->>A: Task 6: Compare
    A-->>O: Comparison complete
    O->>W: Task 7: Create report
    W-->>O: Report complete
    O->>O: Synthesize results
    O->>U: Final response + PDF

sequenceDiagram
    participant U as User
    participant O as Overlord
    participant R as Researcher
    participant A as Analyst
    participant W as Writer

    U->>O: Complex request
    O->>O: Score complexity: 9
    O->>O: Create workflow
    O->>R: Tasks 1-5 (parallel)
    R-->>O: All research complete
    O->>A: Task 6: Compare
    A-->>O: Comparison complete
    O->>W: Task 7: Create report
    W-->>O: Report complete
    O->>O: Synthesize results
    O->>U: Final response + PDF

Error Handling

What happens when a task fails?

Task 1: ✅ Success
Task 2: ❌ Failed (API timeout)
Task 3: ✅ Success
         ↓
Overlord:
  - Retry Task 2 (up to 3 times)
  - If still fails, mark as partial failure
  - Continue with available results
         ↓
Task 4: Executes with results from 1 and 3
         ↓
Response: "I completed the analysis, but encountered an issue
           accessing one data source. Here are the results..."

Configuration

Enable Workflows

overlord:
  workflow:
    auto_decomposition: true      # Enable automatic workflows
    complexity_threshold: 7.0     # Score to trigger workflow (default: 7.0)

Workflow Settings

overlord:
  workflow:
    max_parallel_tasks: 5         # Max concurrent tasks (default: 5)
    parallel_execution: true      # Enable parallel task execution
    timeouts:
      task_timeout: 300           # Seconds per task (default: 300)
      workflow_timeout: 3600      # Max duration for entire workflow (default: 3600)
    retry:
      max_attempts: 3             # Max retry attempts (default: 3)

Disable for Testing

overlord:
  workflow:
    auto_decomposition: false     # Force single-agent mode

Task Types

Sequential Tasks

Tasks that must run in order:

Task 1: Fetch data
         ↓
Task 2: Process data (needs Task 1 output)
         ↓
Task 3: Generate report (needs Task 2 output)

Parallel Tasks

Tasks that can run simultaneously:

    ┌─ Task 1: Research A
    ├─ Task 2: Research B
    └─ Task 3: Research C
         ↓
    All complete → Task 4: Compare

Conditional Tasks

Tasks that depend on previous results:

Task 1: Check if user has credentials
         ↓
    ├─ Yes → Task 2A: Use existing
    └─ No  → Task 2B: Prompt for credentials

When Workflows Trigger

Automatic triggers:

✅ Multiple distinct steps mentioned
✅ Different domains/skills required
✅ Research + creation/analysis
✅ Multiple output artifacts
✅ Time-consuming operations

Examples:

✅ "Research and write blog post" → Workflow
✅ "Analyze codebase and create report" → Workflow
✅ "Search docs, find answer, create ticket" → Workflow
✅ "Compare 3 products and recommend best" → Workflow

❌ "What's the weather?" → Single agent
❌ "Explain this code" → Single agent
❌ "Search for information" → Single agent

Benefits

1. Automatic Orchestration

No manual coordination needed - MUXI handles everything.

2. Parallel Execution

Independent tasks run simultaneously, reducing total time:

Sequential: 5 tasks × 10s = 50s total
Parallel:   5 tasks → 10s total (all at once)

3. Intelligent Routing

Each task goes to the most capable agent.

4. Error Recovery

Failed tasks retry automatically, partial results still useful.

5. Consistent Results

Same workflow structure for similar requests.

Comparison with SOPs

Feature	Workflows	SOPs
Trigger	Automatic (complexity score)	Manual (keyword match)
Definition	AI-generated on-the-fly	Pre-defined template
Flexibility	Adapts to request	Fixed structure
Use case	One-off complex tasks	Repeatable processes

When to use workflows:

New, complex requests
Ad-hoc analysis
Creative projects

When to use SOPs:

Repeatable processes
Standardized workflows
Compliance requirements

Advanced: Workflow Routing

Configure how tasks are routed to agents:

overlord:
  workflow:
    routing_strategy: "capability_based"  # Default: match task to agent capabilities
    # Other strategies: load_balanced, priority_based, round_robin, specialized
    enable_agent_affinity: true  # Prefer agents that succeeded with similar tasks

Performance Optimization

Adjust Timeouts

overlord:
  workflow:
    timeouts:
      task_timeout: 600          # Longer timeout for complex tasks
      workflow_timeout: 7200     # Max duration for entire workflow (2 hours)

Control Parallelism

overlord:
  workflow:
    parallel_execution: true     # Enable parallel execution
    max_parallel_tasks: 3        # Reduce if hitting rate limits (default: 5)

Configure Retries

overlord:
  workflow:
    retry:
      max_attempts: 5            # More retries for flaky operations (default: 3)
      initial_delay: 1.0         # Seconds before first retry (default: 1.0)
      max_delay: 60.0            # Max retry delay (default: 60.0)
      backoff_factor: 2.0        # Exponential backoff multiplier (default: 2.0)

Debugging Workflows

Enable detailed logging:

Configure logging streams to capture workflow details (see Logging Configuration for more options).

Output:

[WORKFLOW] Created workflow for request req_abc123
[WORKFLOW]   Task 1: research (researcher) - no dependencies
[WORKFLOW]   Task 2: research (researcher) - no dependencies
[WORKFLOW]   Task 3: analyze (analyst) - depends on [1, 2]
[WORKFLOW] Executing Task 1 and Task 2 in parallel
[WORKFLOW] Task 1 completed in 3.2s
[WORKFLOW] Task 2 completed in 4.1s
[WORKFLOW] Executing Task 3 (dependencies met)
[WORKFLOW] Task 3 completed in 2.8s
[WORKFLOW] Workflow completed in 7.3s total

Best Practices

DO:

✅ Let MUXI handle decomposition automatically
✅ Set reasonable timeouts (balance speed vs completion)
✅ Monitor workflow performance in production
✅ Adjust complexity threshold based on your use case
✅ Use default values unless you have specific performance needs

DON'T:

❌ Disable auto_decomposition in production (only for testing)
❌ Set task_timeout too low (causes premature failures)
❌ Set max_parallel_tasks too high (can hit API rate limits)
❌ Manually orchestrate tasks (MUXI does this automatically)

Learn More

Agent Formation Schema - Official formation schema specification
The Overlord - How orchestration works
Human-in-the-Loop - Plan approval workflow
Workflow Configuration Reference - Complete workflow settings
Request Lifecycle - See workflows in action

We use cookies