Parallel Execution - Automagik Suite Documentation

Overview

Parallel execution allows running multiple task attempts simultaneously, dramatically reducing total execution time. Perfect for comparing approaches or working on independent tasks.

Speed boost: Run 3 attempts in parallel = 3x faster than sequential execution!

Why Parallel Execution?

Sequential (Slow)

Task 1 (Claude):  [====] 5 min
Task 2 (Gemini):        [===] 3 min
Task 3 (GPT-4):              [====] 4 min
Total time: 12 minutes

Parallel (Fast)

Task 1 (Claude):  [====] 5 min
Task 2 (Gemini):  [===]  3 min
Task 3 (GPT-4):   [====] 4 min
Total time: 5 minutes (limited by slowest)

Result: 2.4x faster!

Basic Parallel Execution

Via CLI

# Run multiple attempts in parallel
forge task create "Add authentication" --llm claude &
forge task fork 1 --llm gemini &
forge task fork 1 --llm gpt-4 &

wait

# All three run simultaneously!

Via SDK

import { ForgeClient } from '@automagik/forge-sdk';

const forge = new ForgeClient();

// Create task
const task = await forge.tasks.create({
  title: 'Add authentication',
  projectId: 'proj_123'
});

// Start 3 attempts in parallel
const attempts = await Promise.all([
  forge.attempts.create({ taskId: task.id, llm: 'claude' }),
  forge.attempts.create({ taskId: task.id, llm: 'gemini' }),
  forge.attempts.create({ taskId: task.id, llm: 'gpt-4' })
]);

console.log('All attempts started!');

Advanced Parallel Patterns

Parallel Task Creation

Create and start multiple tasks at once:

const tasks = await Promise.all([
  forge.tasks.create({
    title: 'Add authentication',
    llm: 'claude'
  }),
  forge.tasks.create({
    title: 'Add dark mode',
    llm: 'gemini'
  }),
  forge.tasks.create({
    title: 'Add search',
    llm: 'cursor'
  })
]);

// Start all tasks
await Promise.all(
  tasks.map(task => forge.tasks.start(task.id))
);

Parallel Comparison

Compare multiple agents on the same task:

async function compareAgents(taskId: string) {
  const agents = ['claude', 'gemini', 'gpt-4', 'cursor'];

  // Create attempts in parallel
  const attempts = await Promise.all(
    agents.map(agent =>
      forge.attempts.create({ taskId, llm: agent })
    )
  );

  // Wait for all to complete
  await Promise.all(
    attempts.map(attempt =>
      forge.attempts.waitForCompletion(attempt.id)
    )
  );

  // Compare results
  return forge.attempts.compare(
    attempts.map(a => a.id)
  );
}

Resource Management

Concurrency Limits

Don’t run too many at once:

import pLimit from 'p-limit';

// Limit to 3 concurrent executions
const limit = pLimit(3);

const tasks = [/* 10 tasks */];

await Promise.all(
  tasks.map(task =>
    limit(() => forge.tasks.start(task.id))
  )
);

// Only 3 run at a time

Rate Limiting

Respect API rate limits:

import pThrottle from 'p-throttle';

// Max 5 requests per second
const throttle = pThrottle({
  limit: 5,
  interval: 1000
});

const throttled = throttle(async (task) => {
  return forge.tasks.create(task);
});

await Promise.all(
  taskData.map(data => throttled(data))
);

Git Worktree Isolation

Each parallel attempt gets its own worktree:

.forge/worktrees/
├── task-1-claude/     ← Isolated workspace
├── task-1-gemini/     ← Isolated workspace
├── task-1-gpt-4/      ← Isolated workspace
└── task-2-cursor/     ← Different task

Benefits:

No conflicts between attempts
Safe parallel modifications
Easy to compare results

Automatic cleanup:

# Forge automatically cleans up after merge
forge task merge 1 --attempt 2

# Or manually
forge worktree cleanup

Monitoring Parallel Execution

Real-Time Dashboard

async function monitorParallelTasks(taskIds: string[]) {
  const interval = setInterval(async () => {
    const statuses = await Promise.all(
      taskIds.map(async id => {
        const task = await forge.tasks.get(id);
        return {
          id,
          title: task.title,
          status: task.status,
          progress: task.attempts[0]?.progress || 0
        };
      })
    );

    console.clear();
    console.table(statuses);

    if (statuses.every(s => s.status === 'completed')) {
      clearInterval(interval);
    }
  }, 1000);
}

Progress Aggregation

async function getOverallProgress(taskIds: string[]) {
  const tasks = await Promise.all(
    taskIds.map(id => forge.tasks.get(id))
  );

  const total = tasks.length;
  const completed = tasks.filter(t => t.status === 'completed').length;
  const inProgress = tasks.filter(t => t.status === 'in_progress').length;

  return {
    total,
    completed,
    inProgress,
    percentage: (completed / total) * 100
  };
}

Error Handling

Graceful Degradation

async function parallelWithFallback(taskId: string) {
  const agents = ['claude', 'gemini', 'gpt-4'];

  const results = await Promise.allSettled(
    agents.map(agent =>
      forge.attempts.create({ taskId, llm: agent })
        .then(attempt => forge.attempts.start(attempt.id))
    )
  );

  // Check which succeeded
  const succeeded = results.filter(r => r.status === 'fulfilled');
  const failed = results.filter(r => r.status === 'rejected');

  console.log(`${succeeded.length} succeeded, ${failed.length} failed`);

  // Use whichever succeeded
  return succeeded.map(r => r.value);
}

Timeout Handling

async function withTimeout<T>(
  promise: Promise<T>,
  timeoutMs: number
): Promise<T> {
  const timeout = new Promise<never>((_, reject) =>
    setTimeout(() => reject(new Error('Timeout')), timeoutMs)
  );

  return Promise.race([promise, timeout]);
}

// Use it
const attempts = await Promise.all([
  withTimeout(
    forge.attempts.create({ taskId, llm: 'claude' }),
    60000  // 1 minute timeout
  ),
  withTimeout(
    forge.attempts.create({ taskId, llm: 'gemini' }),
    60000
  )
]);

Cost Optimization

Smart Parallel Strategy

async function costOptimizedParallel(taskId: string) {
  // Step 1: Quick exploration with cheap models
  const cheapAttempts = await Promise.all([
    forge.attempts.create({ taskId, llm: 'gemini' }),  // Free!
    forge.attempts.create({ taskId, llm: 'claude-haiku' })  // Cheap
  ]);

  await Promise.all(
    cheapAttempts.map(a => forge.attempts.waitForCompletion(a.id))
  );

  // Step 2: Review cheap results
  const cheapResults = await forge.attempts.compare(
    cheapAttempts.map(a => a.id)
  );

  // Step 3: Only use expensive model if needed
  if (cheapResults.quality < 0.8) {
    const expensiveAttempt = await forge.attempts.create({
      taskId,
      llm: 'claude-sonnet'
    });

    await forge.attempts.waitForCompletion(expensiveAttempt.id);
  }

  // Saved money by not always using expensive model!
}

Batch Operations

Bulk Task Processing

async function processBulkTasks(tasks: TaskInput[]) {
  // Create all tasks in parallel
  const created = await Promise.all(
    tasks.map(task => forge.tasks.create(task))
  );

  // Start all tasks in parallel (with concurrency limit)
  const limit = pLimit(5);
  await Promise.all(
    created.map(task =>
      limit(() => forge.tasks.start(task.id))
    )
  );

  // Wait for all to complete
  await Promise.all(
    created.map(task =>
      forge.tasks.waitForCompletion(task.id)
    )
  );

  return created;
}

Parallel Template Execution

async function parallelTemplateRun(
  templateName: string,
  configs: TemplateConfig[]
) {
  // Use template with different configs in parallel
  const results = await Promise.all(
    configs.map(config =>
      forge.templates.use(templateName, config)
    )
  );

  return results;
}

// Example: Test API with different configurations
await parallelTemplateRun('api-endpoint', [
  { resourceName: 'user', methods: ['GET', 'POST'] },
  { resourceName: 'product', methods: ['GET', 'POST', 'PUT'] },
  { resourceName: 'order', methods: ['GET', 'POST', 'DELETE'] }
]);

Real-World Examples

Example 1: A/B/C Testing

async function abcTest(featureName: string) {
  // Try 3 different implementations simultaneously
  const approaches = [
    { llm: 'claude', agent: 'performance-optimizer' },
    { llm: 'gemini', agent: 'simple-clean' },
    { llm: 'gpt-4', agent: 'comprehensive' }
  ];

  const task = await forge.tasks.create({
    title: `Implement ${featureName}`,
    description: 'Try different approaches'
  });

  // Run all approaches in parallel
  const attempts = await Promise.all(
    approaches.map(({ llm, agent }) =>
      forge.attempts.create({
        taskId: task.id,
        llm,
        agent
      })
    )
  );

  // Wait for completion
  await Promise.all(
    attempts.map(a => forge.attempts.waitForCompletion(a.id))
  );

  // Compare results
  const comparison = await forge.attempts.compare(
    attempts.map(a => a.id)
  );

  // Pick winner
  return comparison.winner;
}

Example 2: Multi-Platform Development

async function multiPlatformDevelopment(feature: string) {
  // Develop for multiple platforms simultaneously
  const platforms = [
    { platform: 'web', llm: 'cursor' },
    { platform: 'mobile', llm: 'claude' },
    { platform: 'desktop', llm: 'gemini' }
  ];

  const tasks = await Promise.all(
    platforms.map(({ platform, llm }) =>
      forge.tasks.create({
        title: `${feature} for ${platform}`,
        llm
      })
    )
  );

  // All platforms developed in parallel!
  await Promise.all(
    tasks.map(task => forge.tasks.start(task.id))
  );

  return tasks;
}

Best Practices

Limit Concurrency

// Good ✅
const limit = pLimit(3);
await Promise.all(
  tasks.map(t => limit(() => runTask(t)))
);

// Bad ❌
await Promise.all(
  tasks.map(t => runTask(t))
);
// Could spawn 100+ parallel tasks!

Handle Failures

// Use Promise.allSettled
const results = await Promise.allSettled([
  attempt1,
  attempt2,
  attempt3
]);

// Check which succeeded
const successful = results
  .filter(r => r.status === 'fulfilled')
  .map(r => r.value);

Monitor Resources

# Check system resources
htop

# Watch Forge processes
forge processes list --watch

# Check disk usage
du -sh .forge/worktrees

Clean Up

// After parallel execution
await forge.worktrees.cleanup();

// Remove failed attempts
await forge.attempts.deleteWhere({
  status: 'failed'
});

Performance Metrics

Track parallel execution performance:

async function measureParallelPerformance() {
  const sequential = await measureSequential();
  const parallel = await measureParallel();

  return {
    sequential: {
      time: sequential.time,
      cost: sequential.cost
    },
    parallel: {
      time: parallel.time,
      cost: parallel.cost
    },
    speedup: sequential.time / parallel.time,
    costRatio: parallel.cost / sequential.cost
  };
}

// Example output:
// {
//   sequential: { time: 720000, cost: 0.45 },
//   parallel: { time: 300000, cost: 0.45 },
//   speedup: 2.4,
//   costRatio: 1.0
// }
// Result: 2.4x faster, same cost!

Next Steps

Specialized Agents

Run different agents in parallel

Task Templates

Templates with parallel tasks

Processes API

Monitor parallel executions

Workflows

Complete workflows with parallelism

Getting Started

Learn

Configuration

Reference

Troubleshooting

Documentation Index

​Overview

​Why Parallel Execution?

​Sequential (Slow)

​Parallel (Fast)

​Basic Parallel Execution

​Via CLI

​Via SDK

​Advanced Parallel Patterns

​Parallel Task Creation

​Parallel Comparison

​Resource Management

​Concurrency Limits

​Rate Limiting

​Git Worktree Isolation

​Monitoring Parallel Execution

​Real-Time Dashboard

​Progress Aggregation

​Error Handling

​Graceful Degradation

​Timeout Handling

​Cost Optimization

​Smart Parallel Strategy

​Batch Operations

​Bulk Task Processing

​Parallel Template Execution

​Real-World Examples

​Example 1: A/B/C Testing

​Example 2: Multi-Platform Development

​Best Practices

Limit Concurrency

Handle Failures

Monitor Resources

Clean Up

​Performance Metrics

​Next Steps

Specialized Agents

Task Templates

Processes API

Workflows

Overview

Why Parallel Execution?

Sequential (Slow)

Parallel (Fast)

Basic Parallel Execution

Via CLI

Via SDK

Advanced Parallel Patterns

Parallel Task Creation

Parallel Comparison

Resource Management

Concurrency Limits

Rate Limiting

Git Worktree Isolation

Monitoring Parallel Execution

Real-Time Dashboard

Progress Aggregation

Error Handling

Graceful Degradation

Timeout Handling

Cost Optimization

Smart Parallel Strategy

Batch Operations

Bulk Task Processing

Parallel Template Execution

Real-World Examples

Example 1: A/B/C Testing

Example 2: Multi-Platform Development

Best Practices

Performance Metrics

Next Steps