Agents

Each agent in Kaizen's pipeline can be individually configured with its own model, settings, and custom prompt.

Viewing Agents

Go to Settings > Agents (/settings/agents) to see all agents. Each entry shows:

• Agent name and role

• Current model

• Whether extended thinking is enabled

• Extra model badges (image, file, audio, video models)

Agent Settings

Click on any agent to configure:

Model

Choose which AI model the agent uses. Any model available on OpenRouter can be selected. Different agents benefit from different models:

• Fast, cheap models (Gemini Flash) work well for routing and planning

• Capable models (Claude Sonnet, GPT-4o) work well for execution

• Thinking models (Claude with thinking) work well for complex code

Specialized Models

Some agents support specialized models for different content types:

Model Type	Purpose
Image Model	Processing and generating images
File Model	Analyzing documents
Audio Model	Transcribing audio
Video Model	Analyzing video content

Extended Thinking

Toggle extended thinking to give the model more reasoning time for complex problems. This is especially useful for the Developer agent when writing complex code.

Timeout

Set the maximum time (10-600 seconds) for a single model call. Increase this for complex tasks that need more processing time.

System Prompt

Each agent has a system prompt that defines its behavior and capabilities. You can view and customize the prompt for each agent. Changes are auto-saved.

Interactive Planning

The Planner agent has a special Interactive Planning toggle. When enabled, the planner proposes its plan to you for approval before execution proceeds. When disabled, planning happens automatically.

Default Agents

Agent	Role	Default Model
Router	Classifies user intent and picks the right path	Claude Sonnet
Planner	Researches the task and builds a step-by-step approach	Gemini Flash
Executor	Does the work using tools, skills, and integrations	Claude Sonnet
Developer	Writes and tests code in the code pipeline	Claude Sonnet (thinking)
Reviewer	Inspects code pipeline output and describes what it sees	Varies by modality
Titler	Generates chat titles	Gemini Flash
Compactor	Compresses long conversation histories	Gemini Flash