Autonomous Agents

Build features, fix bugs, and refactor code with AI agents that work autonomously.

Overview

Autonomous agents go beyond chat—they take action. Give them a task, and they’ll plan, code, test, and iterate until it’s done. You stay in control with approval checkpoints and configurable permissions.

One-Shot Tasks

The fastest way to build something from scratch.

How It Works

1. Describe your goal - “Build a REST API for user management with Express.js and MongoDB”
2. Set completion criteria - Define what “done” looks like:
   • All tests pass
   • API documentation generated
   • Docker compose file included
3. Configure execution:
   • Sandbox: Local (fastest) or Isolated (secure)
   • Provider: Anthropic, OpenAI, etc.
   • Model: Claude, GPT-4o, etc.
   • Backend: Auto, TurboLight, DeepCode
   • Iterations: 10 (quick) to 100 (thorough)
4. Start the agent - Watch it work or come back later

Example Tasks

Build a feature:

Build a user authentication system with:
- JWT tokens
- Password hashing
- Login/logout endpoints
- Protected route middleware

Refactor code:

Refactor the payment module to use the strategy pattern.
Extract each payment provider into its own class.
Add comprehensive tests.

Fix a bug:

The checkout flow fails when users have items
with special characters in the name.
Find and fix the root cause.
Add regression tests.

Agent Backends

Different backends optimize for different scenarios:

AI Councils

When you need multiple perspectives, convene a council.

How Councils Work

1. You pose a question or task
2. Multiple AI agents discuss it
3. They deliberate and debate
4. They reach consensus (or report disagreement)
5. You get a recommendation with reasoning

Preset Councils

Technical Review (4 members)

• Architect, Security, Performance, Quality
• Multi-perspective code review
• Catches issues a single reviewer might miss

Quick Review (1 member)

• Fast feedback for small changes
• Good for typos, minor refactors

Architecture Decision (3 members)

• System design choices
• “Should we use microservices or monolith?”

Product Council (3 members)

• Feature prioritization
• Trade-off analysis

Model Showdown (3 models)

• Compare responses from different AI models
• Find the best model for your task

Code Duel (2 members)

• Two agents implement the same feature
• Compare approaches

Supervised Team (3 members)

• Senior model oversees junior models
• Mentored development with quality control

Review Board (2 members)

• Thorough code review
• Detailed feedback

Council Configuration

Creating Custom Councils

1. Click + in the Councils panel
2. Set council name and description
3. Add members from:
   • Native LLM - Raw model access
   • Agent Engine - Tool-using agents
   • Amplifier Agents - Specialized experts
   • CLI Agent - External AI CLIs
4. Configure deliberation settings
5. Save as template for reuse

Agent Settings

Fine-tune how agents behave across all tasks.

General Settings

Behavior Settings

Permission Profiles

Strict

• Approval required for all file changes
• No command execution without confirmation
• Best for: Production code, sensitive projects

Balanced (Recommended)

• Auto-approve read operations
• Approve writes and commands
• Best for: Most development work

Permissive

• Auto-approve most safe operations
• Only confirm destructive actions
• Best for: Prototyping, personal projects

Background Agents

Run agents in the background while you work.

Starting Background Tasks

1. Configure your task as usual
2. Enable “Run in Background”
3. Continue working on other things
4. Get notified when complete

Managing Background Agents

• View running agents in the status bar
• Check progress anytime
• Cancel if needed
• Review results when done

Best Practices

Writing Good Task Descriptions

Be specific:

Good: "Add pagination to the /users endpoint with
      page and limit query params, default 20 per page"

Vague: "Add pagination"

Include context:

Good: "This is a Next.js 14 app using Prisma.
      Add a dark mode toggle to the settings page."

Missing context: "Add dark mode"

Define success:

Good: "Refactor auth to use refresh tokens.
      Criteria: All existing tests pass,
      new tests for refresh flow, docs updated."

No criteria: "Refactor auth"

Choosing Iterations

When to Use Councils

• Architecture decisions - Get multiple perspectives
• Code review - Catch more issues
• Trade-off analysis - Explore options
• Learning - See different approaches