V
Viban
Get Started

AI Agents

Viban supports multiple AI agents for autonomous task execution. Learn how each agent works and when to use them.

Supported Agents

Claude Code (Default)

Claude Code is Anthropic's autonomous coding agent, designed specifically for software development tasks.

Strengths:

  • Excellent code understanding and generation
  • Strong reasoning capabilities
  • Built-in tool use (file editing, terminal, etc.)
  • Long context windows
  • Full project context awareness

Best For:

  • Complex feature implementation
  • Refactoring and code cleanup
  • Bug fixing with context
  • Documentation generation

Configuration:

# Install
npm install -g @anthropic-ai/claude-code

# Authenticate
claude login

Gemini CLI

Google's Gemini model accessible through the CLI for code tasks.

Strengths:

  • Multimodal understanding (can process images)
  • Large context window
  • Fast inference

Best For:

  • Tasks involving images/diagrams
  • Documentation from mockups
  • Multi-file understanding

Configuration:

# Install Gemini CLI
# Follow Google's official installation guide

Codex

OpenAI's Codex executor for code generation tasks.

Strengths:

  • Strong code completion capabilities
  • Wide language support

OpenCode

OpenCode agent for autonomous development.

Cursor Agent

Cursor AI agent for IDE-integrated development tasks.

Agent Selection

When starting a task, you can choose which agent to use:

  1. Open task details
  2. Select agent from dropdown
  3. Drag to "In Progress"

Automatic Selection

Viban can automatically select the best agent based on:

  • Task complexity
  • Required capabilities
  • Agent availability

Execution Model

Isolated Environments

Each task runs in an isolated git worktree:

/workspaces/
  /task-abc123/
    /.git  (worktree reference)
    /src
    /...

This ensures:

  • No conflicts between parallel tasks
  • Clean rollback if needed
  • Easy code review

Real-time Streaming

Output streams to the UI in real-time:

  1. AI agent writes to stdout
  2. Captured by execution worker
  3. Broadcast via Phoenix channels
  4. Rendered in task detail view

Error Handling

If an agent encounters an error:

  1. Error is logged to execution history
  2. Task remains in "In Progress"
  3. You can retry or cancel
  4. Worktree is preserved for debugging

Resource Management

Concurrent Execution

Configure maximum concurrent tasks:

# config/config.exs
config :viban, :execution,
  max_concurrent_tasks: 3

Timeouts

Default timeout is 30 minutes. Configure per-task:

config :viban, :execution,
  default_timeout: :timer.minutes(30)

Monitoring

Execution Logs

View detailed logs in the task detail panel:

  • Command executed
  • Start/end timestamps
  • Exit codes
  • Output history

Metrics

Track agent performance:

  • Success rate
  • Average completion time
  • Token usage (where applicable)