Agents

Agents are the core workers in Mule. Each agent is configured with an AI Provider, Model, and a system prompt that defines its behavior.

Agent Configuration link

When creating an agent, you configure:

• Name: A descriptive identifier for the agent
• Description: What this agent does (optional)
• AI Provider: The configured AI provider to use
• Model: The specific model from that provider (format: provider/model-id, e.g., anthropic/claude-sonnet-4.5)
• System Prompt: Instructions that define the agent’s behavior and capabilities
• Tools: External capabilities the agent can use (filesystem, bash, HTTP, database, memory)
• Skills: pi agent skills that extend the agent’s capabilities
• PI Config: Advanced settings for the PI runtime (thinking level, etc.)

PI Configuration link

The PI Config allows you to fine-tune how the agent processes tasks. This is passed as a JSON object when creating or updating an agent:

{
  "thinking_level": "medium",
  "skills": ["skill-id-1", "skill-id-2"],
  "tools": ["read", "write", "edit", "bash", "grep", "find"],
  "extensions": ["extension-name"],
  "working_dir": "/path/to/default/directory"
}

Thinking Levels link

The thinking_level setting controls how much reasoning the agent applies before responding:

Higher thinking levels result in more thoughtful responses but take longer to generate. Adjust based on task complexity.

Additional Configuration Options link

The PI Config supports several additional options:

Note: There are two ways to configure skills for agents:

1. PI Config skills array: Lists pi skill identifiers (e.g., search-cli, actual-budget) that the pi runtime will use
2. Database skill assignment: Assigns Mule-managed skills stored in the database via /api/v1/agents/{id}/skills Both can be used together - PI Config skills are used by the pi runtime, while database skills are managed by Mule.

Available Tools link

When configuring tools in PI Config, you can specify any of these available tools:

System Prompt link

The system prompt defines how the agent behaves. The system prompt is sent as-is to the AI model, so you can include any instructions you want the agent to follow.

For best results with code-related tasks, include:

• Clear instructions on what the agent should do
• Examples of expected input and output
• Any constraints or requirements
• References to relevant files or patterns

Example system prompt:

You are a code review assistant. Your task is to:
1. Read the provided code changes
2. Identify potential bugs, security issues, or style problems
3. Suggest improvements in a clear, actionable format

When reviewing, consider:
- Code clarity and maintainability
- Error handling
- Performance implications
- Security best practices

Tools link

Agents can be assigned tools to interact with external systems. For complete documentation on available tools and configuration, see the Tools documentation.

• filesystem: Read and write files on the system
• bash: Execute shell commands
• http: Make HTTP requests to external APIs
• database: Query databases directly
• memory: Store and retrieve semantic memories using vector embeddings

Memory Tool Configuration link

When using the memory tool, the agent can store and retrieve information using semantic search. The memory configuration includes:

• Database URL: PostgreSQL database for vector storage
• Embedding Provider: AI provider for generating embeddings
• Embedding Model: Specific embedding model to use
• Embedding Dimensions: Vector size for the embedding model
• Default TTL: Time-to-live for stored memories (in seconds)
• Default Top-K: Number of similar memories to retrieve

Skills link

Skills extend agents with specialized capabilities using pi agent skills. Skills are stored at a specific path and can be enabled or disabled per skill.

To assign skills to an agent:

1. Create the skill with a name, description, and path
2. Assign the skill to the agent via the agent’s skills configuration

Skills are executed by the pi runtime and can provide additional tools, knowledge, or capabilities to the agent.

Execution Settings link

Working Directory link

Agents can be configured to operate within a specific working directory. This is useful for:

• Repository-specific operations
• Keeping file operations within a project structure
• Maintaining consistent paths across tool executions

Timeout link

Agents have a default timeout of 5 minutes for execution. If an agent takes longer, the operation will be aborted. Adjust the timeout based on the complexity of expected tasks.

Image Support link

Agents can process images in prompts when the AI provider supports vision capabilities. This enables:

• Analyzing screenshots or diagrams
• Reading visual code representations
• Processing images as part of the task context

Validation Functions link

After an agent generates code, validation can be performed through WASM modules. This allows you to:

• Run linters (gofmt, golangci-lint, ESLint, etc.)
• Execute tests (go test, pytest, etc.)
• Check code formatting
• Validate dependencies
• Run custom build checks

Validation modules receive the generated code as input and return pass/fail results. If validation fails, Mule can make additional attempts to fix issues.

Multi-Agent Workflows link

Multiple agents can be used together in workflows, with each agent handling different responsibilities:

• Analysis Agent: Reviews issues and plans code changes
• Generation Agent: Writes the actual code
• Review Agent: Validates and improves generated code
• Commit Agent: Creates appropriate commit messages

See the Workflows documentation for details on chaining agents together in pipelines.

Workflow Integration link

Agents are typically used within workflows, where each workflow step can reference a specific agent. This allows you to:

• Chain multiple agents sequentially
• Pass output from one agent as input to another
• Include validation steps between agent executions
• Create sophisticated multi-agent pipelines for complex tasks