Tools

Overview

Tools are the mechanism through which OrbitAI agents interact with the external world and perform specialized operations. They enable agents to go beyond language processing to take concrete actions, access external resources, and manipulate data.

Extensible

Built-in tools and custom tool creation support

Type-Safe

JSON Schema validation ensures correct usage

Async-Ready

Fully asynchronous execution with Swift Concurrency

Traceable

Comprehensive metrics and execution tracking

Intelligent

LLM-driven automatic tool selection

Composable

Combine tools for complex workflows

Key Capabilities

Automatic Tool Selection

Agents automatically determine which tools to use based on task requirements, available tools, and context. The LLM reasons about tool usage and orchestrates multi-tool workflows.

Schema Validation

Every tool defines its input parameters using JSON Schema, ensuring type safety and preventing invalid tool usage before execution.

Execution Tracking

Tool execution is monitored with detailed metrics including execution time, success status, input/output sizes, and error information.

Error Handling

Built-in error handling mechanisms ensure graceful failures with detailed error messages and recovery strategies.

Tool Architecture

Tool System
    ├── ToolsHandler (Singleton)
    │   ├── Tool Registry
    │   ├── Registration/Lookup
    │   └── Lifecycle Management
    │
    ├── BaseTool (Base Class)
    │   ├── Name & Description
    │   ├── Parameters Schema
    │   └── Execute Method
    │
    ├── Built-in Tools
    │   ├── File Operations
    │   ├── Web Access
    │   ├── Apple Platform APIs
    │   └── Data Processing
    │
    ├── Custom Tools
    │   ├── User-Defined Logic
    │   ├── External API Integration
    │   └── Domain-Specific Tools
    │
    └── Tool Execution
        ├── Parameter Validation
        ├── Async Execution
        ├── Result Handling
        └── Metrics Collection

Built-in Tools

OrbitAI includes a comprehensive suite of built-in tools for common operations:

File & Data Operations

file_reader

Description: Read and process file contents Parameters: path (string), encoding (optional) Returns: File contents as string or data

file_writer

Description: Write or update file contents Parameters: path (string), content (string), mode (append/overwrite) Returns: Success status and file path

directory_list

Description: List directory contents with filtering Parameters: path (string), recursive (boolean), filter (regex) Returns: Array of file/directory information

csv_processor

Description: Parse and manipulate CSV data Parameters: file_path (string), operation (read/write/transform) Returns: Structured data or success status

data_analyzer

Description: Analyze datasets with statistical operations Parameters: data (array), operations (mean/median/std/etc) Returns: Statistical analysis results

Web & Network Tools

web_search

Description: Search the internet for current information Parameters: query (string), num_results (integer, 1-10) Returns: Array of search results with URLs and snippets

web_scraping

Description: Extract content from web pages Parameters: url (string), selectors (CSS/XPath) Returns: Extracted content and metadata

api_caller

Description: Make HTTP requests to external APIs Parameters: url, method, headers, body, auth Returns: Response data and status

geocoding

Description: Convert addresses to coordinates and vice versa Parameters: query (string), type (forward/reverse) Returns: Location data with coordinates

Apple Platform Integration

apple_calendar

Description: Interact with Apple Calendar (EventKit) Parameters: operation (read/create/update/delete), event_data Returns: Calendar events or success status

apple_reminders

Description: Manage Apple Reminders Parameters: operation, reminder_data, list_name Returns: Reminders or success status

local_notifications

Description: Schedule and manage local notifications Parameters: title, body, trigger, identifier Returns: Notification identifier and delivery status

core_location

Description: Access device location services Parameters: accuracy (best/kilometer/etc), continuous (boolean) Returns: Location coordinates and metadata

weather_kit

Description: Retrieve weather information using WeatherKit Parameters: location (coordinates/place), forecast_type Returns: Weather data and forecasts

Computation Tools

calculator

Description: Perform mathematical calculations Parameters: expression (string), precision (integer) Returns: Calculation result

code_executor

Description: Execute code in sandboxed environment Parameters: code (string), language (python/swift/js) Returns: Execution output and status

chart_generator

Description: Create data visualizations Parameters: data, chart_type, styling Returns: Chart image or data URL

Built-in tools are automatically registered with ToolsHandler when OrbitAI initializes. No manual registration is required.

Tool Assignment

Tools can be assigned at both the agent and task levels, providing flexibility in tool availability and usage.

Agent-Level Tools

Assign tools to an agent to define its general capabilities across all tasks:

let researchAgent = Agent(
    role: "Research Analyst",
    purpose: "Conduct comprehensive market research and analysis",
    context: """
    Expert researcher with access to web resources,
    data analysis tools, and visualization capabilities.
    """,
    tools: [
        "web_search",
        "web_scraping",
        "data_analyzer",
        "chart_generator",
        "file_writer"
    ]
)

When to use: When tools are fundamental to the agent’s role and will be used across multiple tasks.

Task-Level Tools

Assign tools to specific tasks for fine-grained control:

let analysisTask = ORTask(
    description: """
    Download the quarterly sales data, analyze trends,
    and generate a visualization report.
    """,
    expectedOutput: "Sales trend report with charts",
    agent: dataAgent,
    tools: [
        "api_caller",          // Fetch sales data
        "csv_processor",       // Process data
        "data_analyzer",       // Analyze trends
        "chart_generator",     // Create visualizations
        "file_writer"          // Save report
    ]
)

When to use: When specific tools are only needed for particular tasks or to restrict tool usage.

Agent Tools vs Task Tools

Tool Resolution
Best Practices
Examples

When both agent and task define tools, OrbitAI uses the following resolution logic:

Effective Tools = Task Tools.isEmpty ? Agent Tools : Task Tools

Key Rules:

If task has tools defined → Use only task tools
If task has no tools → Use agent tools
Task tools override agent tools (not merge)
Empty tool list [] means no tools available

Task tools completely override agent tools—they do not merge. If you need agent tools plus additional task tools, explicitly list all required tools in the task definition.

Creating Custom Tools

Custom tools enable integration with proprietary systems, external APIs, and domain-specific functionality.

Basic Custom Tool

Create a custom tool by extending BaseTool:

Define Tool Class

import OrbitAI

final class DatabaseQueryTool: BaseTool {
    override var name: String {
        "database_query"
    }

    override var description: String {
        """
        Query the company database for information.
        Use this when you need to retrieve records,
        check inventory, or analyze stored data.
        """
    }
}

Define Parameters Schema

override var parametersSchema: JSONSchema {
    JSONSchema(
        type: .object,
        properties: [
            "query": JSONSchema(
                type: .string,
                description: "SQL query to execute"
            ),
            "limit": JSONSchema(
                type: .integer,
                description: "Maximum rows to return (1-1000)"
            ),
            "timeout": JSONSchema(
                type: .number,
                description: "Query timeout in seconds"
            )
        ],
        required: ["query"]
    )
}

Implement Execute Method

override func execute(
    with parameters: Metadata
) async throws -> ToolResult {
    // Validate parameters
    guard let query = parameters["query"]?.stringValue else {
        throw OrbitAIError.invalidToolParameters(
            "Missing required parameter: query"
        )
    }

    // Apply defaults
    let limit = parameters["limit"]?.intValue ?? 100
    let timeout = parameters["timeout"]?.doubleValue ?? 30.0

    // Validate constraints
    guard (1...1000).contains(limit) else {
        throw OrbitAIError.invalidToolParameters(
            "Limit must be between 1 and 1000"
        )
    }

    // Execute operation
    let results = try await database.execute(
        query,
        limit: limit,
        timeout: timeout
    )

    // Format results
    var resultData = Metadata()
    resultData["rows"] = .array(results.map { .dictionary($0) })
    resultData["count"] = .int(results.count)
    resultData["query"] = .string(query)

    return ToolResult(success: true, data: resultData)
}

// Register globally
let toolsHandler = ToolsHandler.shared
await toolsHandler.registerTool(DatabaseQueryTool())

// Now available to all agents
let agent = Agent(
    role: "Data Analyst",
    purpose: "Query and analyze company data",
    context: "Expert in SQL and data analysis",
    tools: ["database_query", "data_analyzer"]
)

Advanced Custom Tool Example

Here’s a more sophisticated tool with error handling, validation, and metrics:

import OrbitAI
import Foundation

final class EmailSenderTool: BaseTool {
    private let emailService: EmailService

    init(emailService: EmailService) {
        self.emailService = emailService
    }

    override var name: String { "send_email" }

    override var description: String {
        """
        Send email messages to recipients with optional attachments.
        Supports both plain text and HTML formatting.
        """
    }

    override var parametersSchema: JSONSchema {
        JSONSchema(
            type: .object,
            properties: [
                "to": JSONSchema(
                    type: .array,
                    items: JSONSchema(type: .string),
                    description: "Recipient email addresses"
                ),
                "subject": JSONSchema(
                    type: .string,
                    description: "Email subject line"
                ),
                "body": JSONSchema(
                    type: .string,
                    description: "Email body content"
                ),
                "format": JSONSchema(
                    type: .string,
                    enum: ["text", "html"],
                    description: "Body format (default: text)"
                ),
                "attachments": JSONSchema(
                    type: .array,
                    items: JSONSchema(type: .string),
                    description: "File paths to attach"
                ),
                "cc": JSONSchema(
                    type: .array,
                    items: JSONSchema(type: .string),
                    description: "CC recipients (optional)"
                ),
                "bcc": JSONSchema(
                    type: .array,
                    items: JSONSchema(type: .string),
                    description: "BCC recipients (optional)"
                )
            ],
            required: ["to", "subject", "body"]
        )
    }

    override func execute(
        with parameters: Metadata
    ) async throws -> ToolResult {
        // Validate required parameters
        guard let recipients = parameters["to"]?.arrayValue?.compactMap({
            $0.stringValue
        }), !recipients.isEmpty else {
            return ToolResult(
                success: false,
                error: "At least one recipient email is required"
            )
        }

        guard let subject = parameters["subject"]?.stringValue,
              !subject.isEmpty else {
            return ToolResult(
                success: false,
                error: "Email subject cannot be empty"
            )
        }

        guard let body = parameters["body"]?.stringValue else {
            return ToolResult(
                success: false,
                error: "Email body is required"
            )
        }

        // Validate email addresses
        for email in recipients {
            guard isValidEmail(email) else {
                return ToolResult(
                    success: false,
                    error: "Invalid email address: \(email)"
                )
            }
        }

        // Parse optional parameters
        let format = parameters["format"]?.stringValue ?? "text"
        let cc = parameters["cc"]?.arrayValue?.compactMap {
            $0.stringValue
        } ?? []
        let bcc = parameters["bcc"]?.arrayValue?.compactMap {
            $0.stringValue
        } ?? []
        let attachmentPaths = parameters["attachments"]?.arrayValue?.compactMap {
            $0.stringValue
        } ?? []

        // Validate attachments exist
        for path in attachmentPaths {
            guard FileManager.default.fileExists(atPath: path) else {
                return ToolResult(
                    success: false,
                    error: "Attachment not found: \(path)"
                )
            }
        }

        // Send email
        do {
            let messageID = try await emailService.send(
                to: recipients,
                cc: cc,
                bcc: bcc,
                subject: subject,
                body: body,
                format: format == "html" ? .html : .text,
                attachments: attachmentPaths
            )

            // Build result
            var resultData = Metadata()
            resultData["message_id"] = .string(messageID)
            resultData["recipients_count"] = .int(recipients.count)
            resultData["sent_at"] = .string(ISO8601DateFormatter().string(from: Date()))

            return ToolResult(success: true, data: resultData)

        } catch {
            return ToolResult(
                success: false,
                error: "Failed to send email: \(error.localizedDescription)"
            )
        }
    }

    private func isValidEmail(_ email: String) -> Bool {
        let emailRegex = #"^[A-Z0-9._%+-]+@[A-Z0-9.-]+\.[A-Z]{2,}$"#
        return email.range(
            of: emailRegex,
            options: [.caseInsensitive, .regularExpression]
        ) != nil
    }
}

Always return detailed error messages in ToolResult when execution fails. This helps agents understand what went wrong and potentially retry with corrected parameters.

Tool Integration

With LLM Function Calling

OrbitAI integrates tools with LLM function calling for intelligent tool usage:

// Tools are automatically converted to LLM function schemas
let webSearchTool = ToolSchema(
    function: FunctionSchema(
        name: "web_search",
        description: "Search the web for current information",
        parameters: JSONSchema(
            type: .object,
            properties: [
                "query": JSONSchema(
                    type: .string,
                    description: "Search query"
                ),
                "num_results": JSONSchema(
                    type: .integer,
                    description: "Number of results (1-10)"
                )
            ],
            required: ["query"]
        )
    )
)

// LLM request with tool
let request = LLMRequest(
    messages: [
        .system("You can search the web for current information."),
        .user("What are the latest AI developments this week?")
    ],
    tools: [webSearchTool]
)

// LLM decides to use the tool
let response = try await llm.generateResponse(for: request)

if case .toolCall(let toolCall) = response.finishReason {
    // Execute the tool
    let tool = await toolsHandler.tool(named: toolCall.name)
    let result = try await tool?.execute(with: toolCall.parameters)
}

Multi-Tool Workflows

Agents automatically orchestrate multi-tool workflows:

let task = ORTask(
    description: """
    Research the top 5 AI companies by market cap,
    analyze their stock performance over the last year,
    create a comparative chart, and save a report.
    """,
    expectedOutput: "Stock performance analysis report with charts",
    tools: [
        "web_search",      // 1. Find top AI companies
        "api_caller",      // 2. Fetch stock data
        "data_analyzer",   // 3. Analyze performance
        "chart_generator", // 4. Create charts
        "file_writer"      // 5. Save report
    ]
)

// Agent automatically:
// 1. Uses web_search to find companies
// 2. Uses api_caller to get stock data for each
// 3. Uses data_analyzer to compute metrics
// 4. Uses chart_generator to visualize
// 5. Uses file_writer to save final report

Tool Chaining

Tools can use outputs from previous tools:

// Example: Scrape → Analyze → Visualize pipeline
let pipeline = ORTask(
    description: """
    Scrape product reviews from the given URLs,
    perform sentiment analysis on the reviews,
    and create a sentiment distribution chart.
    """,
    expectedOutput: "Sentiment analysis chart",
    context: """
    URLs: https://example.com/product/reviews
    Analyze positive, negative, and neutral sentiments.
    """,
    tools: [
        "web_scraping",    // Output: Review texts
        "data_analyzer",   // Input: Review texts → Output: Sentiments
        "chart_generator"  // Input: Sentiments → Output: Chart
    ]
)

The agent automatically chains tool outputs as inputs to subsequent tools based on the LLM’s reasoning about the task flow.

Tool Execution

Execution Lifecycle

Tool Execution Lifecycle
    ├── 1. Tool Selection
    │   └── LLM chooses tool based on task
    │
    ├── 2. Parameter Preparation
    │   ├── LLM generates parameters
    │   └── Schema validation
    │
    ├── 3. Pre-Execution
    │   ├── Check tool availability
    │   └── Validate parameters against schema
    │
    ├── 4. Execution
    │   ├── Call tool's execute() method
    │   ├── Async operation
    │   └── Monitor progress
    │
    ├── 5. Result Handling
    │   ├── Parse ToolResult
    │   ├── Check success status
    │   └── Extract data or error
    │
    └── 6. Post-Execution
        ├── Record metrics
        ├── Update agent memory
        └── Continue task or return result

Execution Model

All tool execution in OrbitAI is asynchronous using Swift Concurrency:

override func execute(
    with parameters: Metadata
) async throws -> ToolResult {
    // Tools can perform long-running operations
    let data = try await performAsyncOperation()

    // Can make multiple async calls
    async let result1 = fetchData()
    async let result2 = processData()

    let combined = try await [result1, result2]

    return ToolResult(success: true, data: formatResults(combined))
}

Timeout Handling

Implement timeouts for long-running tools:

override func execute(
    with parameters: Metadata
) async throws -> ToolResult {
    let timeout = parameters["timeout"]?.doubleValue ?? 30.0

    do {
        return try await withThrowingTaskGroup(of: ToolResult.self) { group in
            // Add main execution task
            group.addTask {
                let result = try await self.performOperation(parameters)
                return ToolResult(success: true, data: result)
            }

            // Add timeout task
            group.addTask {
                try await Task.sleep(nanoseconds: UInt64(timeout * 1_000_000_000))
                throw ToolError.timeout
            }

            // Return first result (execution or timeout)
            let result = try await group.next()!
            group.cancelAll()
            return result
        }
    } catch is ToolError {
        return ToolResult(
            success: false,
            error: "Operation timed out after \(timeout)s"
        )
    }
}

Advanced Tool Setup

Tool Registry & Management

The ToolsHandler singleton manages all tools:

let toolsHandler = ToolsHandler.shared

// Register tools
await toolsHandler.registerTool(CustomTool())
await toolsHandler.registerTool(AnotherTool())

// Check tool availability
let isAvailable = await toolsHandler.isToolAvailable(named: "custom_tool")

// Get tool instance
if let tool = await toolsHandler.tool(named: "custom_tool") {
    let result = try await tool.execute(with: params)
}

// List all registered tools
let allTools = await toolsHandler.registeredToolNames()
print("Available tools: \(allTools)")

Enabling/Disabling Tools

Control tool availability dynamically:

final class ConditionalTool: BaseTool {
    private var isEnabled: Bool = true

    override var name: String { "conditional_tool" }

    func enable() {
        isEnabled = true
    }

    func disable() {
        isEnabled = false
    }

    override func execute(
        with parameters: Metadata
    ) async throws -> ToolResult {
        guard isEnabled else {
            return ToolResult(
                success: false,
                error: "Tool is currently disabled"
            )
        }

        // Normal execution
        return try await performOperation(parameters)
    }
}

// Usage
let conditionalTool = ConditionalTool()
await toolsHandler.registerTool(conditionalTool)

// Disable during maintenance
conditionalTool.disable()

// Re-enable when ready
conditionalTool.enable()

Tool Introspection

Tools can be introspected for debugging and documentation:

final class IntrospectableAgent {
    let agent: Agent

    func inspectTools() async -> [ToolInfo] {
        var toolInfos: [ToolInfo] = []

        for toolName in agent.tools {
            if let tool = await ToolsHandler.shared.tool(named: toolName) {
                let info = ToolInfo(
                    name: tool.name,
                    description: tool.description,
                    schema: tool.parametersSchema,
                    requiredParams: extractRequiredParams(from: tool),
                    optionalParams: extractOptionalParams(from: tool)
                )
                toolInfos.append(info)
            }
        }

        return toolInfos
    }

    private func extractRequiredParams(from tool: BaseTool) -> [String] {
        guard case .object(_, _, let required) = tool.parametersSchema else {
            return []
        }
        return required ?? []
    }

    private func extractOptionalParams(from tool: BaseTool) -> [String] {
        guard case .object(let properties, _, let required) = tool.parametersSchema else {
            return []
        }
        let requiredSet = Set(required ?? [])
        return Array(properties.keys.filter { !requiredSet.contains($0) })
    }
}

// Usage
let inspector = IntrospectableAgent(agent: myAgent)
let tools = await inspector.inspectTools()

for tool in tools {
    print("Tool: \(tool.name)")
    print("  Description: \(tool.description)")
    print("  Required: \(tool.requiredParams)")
    print("  Optional: \(tool.optionalParams)")
}

Tool Versioning

Support multiple versions of the same tool:

final class WebSearchToolV1: BaseTool {
    override var name: String { "web_search" }
    // V1 implementation
}

final class WebSearchToolV2: BaseTool {
    override var name: String { "web_search_v2" }
    // V2 implementation with enhanced features
}

// Register both versions
await toolsHandler.registerTool(WebSearchToolV1())
await toolsHandler.registerTool(WebSearchToolV2())

// Use specific version in agent
let modernAgent = Agent(
    role: "Researcher",
    tools: ["web_search_v2"]  // Use latest version
)

let legacyAgent = Agent(
    role: "Legacy Researcher",
    tools: ["web_search"]  // Use V1 for compatibility
)

Usage and Metrics

Tool Usage Tracking

OrbitAI automatically tracks tool usage with detailed metrics:

public struct ToolUsage: Codable, Sendable {
    public let toolName: String           // Name of the tool used
    public let executionTime: TimeInterval // Execution duration
    public let success: Bool              // Execution success status
    public let inputSize: Int             // Size of input parameters
    public let outputSize: Int            // Size of output data
}

Accessing Tool Metrics

// From TaskOutput after task completion
let output = try await orbit.run()

print("Task completed with \(output.toolsUsed.count) tool executions")

for toolUsage in output.toolsUsed {
    print("""
    Tool: \(toolUsage.toolName)
      Execution Time: \(String(format: "%.2f", toolUsage.executionTime))s
      Success: \(toolUsage.success ? "✓" : "✗")
      Input Size: \(toolUsage.inputSize) bytes
      Output Size: \(toolUsage.outputSize) bytes
    """)
}

// Aggregate metrics
let totalExecutionTime = output.toolsUsed.reduce(0) { $0 + $1.executionTime }
let successRate = Double(output.toolsUsed.filter { $0.success }.count) /
                  Double(output.toolsUsed.count)

print("Total tool execution time: \(totalExecutionTime)s")
print("Success rate: \(Int(successRate * 100))%")

Performance Monitoring

Create custom monitoring for tool performance:

final class MonitoredTool: BaseTool {
    private var executionHistory: [ExecutionRecord] = []

    struct ExecutionRecord {
        let timestamp: Date
        let parameters: Metadata
        let success: Bool
        let duration: TimeInterval
        let error: String?
    }

    override func execute(
        with parameters: Metadata
    ) async throws -> ToolResult {
        let startTime = Date()
        var record = ExecutionRecord(
            timestamp: startTime,
            parameters: parameters,
            success: false,
            duration: 0,
            error: nil
        )

        do {
            let result = try await performOperation(parameters)

            record = ExecutionRecord(
                timestamp: startTime,
                parameters: parameters,
                success: true,
                duration: Date().timeIntervalSince(startTime),
                error: nil
            )

            executionHistory.append(record)
            return result

        } catch {
            record = ExecutionRecord(
                timestamp: startTime,
                parameters: parameters,
                success: false,
                duration: Date().timeIntervalSince(startTime),
                error: error.localizedDescription
            )

            executionHistory.append(record)
            throw error
        }
    }

    func getPerformanceStats() -> (avgDuration: TimeInterval, successRate: Double) {
        guard !executionHistory.isEmpty else {
            return (0, 0)
        }

        let avgDuration = executionHistory.reduce(0.0) { $0 + $1.duration } /
                         Double(executionHistory.count)
        let successCount = executionHistory.filter { $0.success }.count
        let successRate = Double(successCount) / Double(executionHistory.count)

        return (avgDuration, successRate)
    }
}

Cost Tracking for External APIs

Track costs for tools that call paid APIs:

final class PaidAPITool: BaseTool {
    private let costPerCall: Decimal
    private var totalCost: Decimal = 0
    private var callCount: Int = 0

    init(costPerCall: Decimal) {
        self.costPerCall = costPerCall
    }

    override func execute(
        with parameters: Metadata
    ) async throws -> ToolResult {
        let result = try await callExternalAPI(parameters)

        // Track costs
        callCount += 1
        totalCost += costPerCall

        // Include cost in result metadata
        var resultData = result.data
        resultData?["cost"] = .string("\(costPerCall)")
        resultData?["total_cost"] = .string("\(totalCost)")
        resultData?["call_count"] = .int(callCount)

        return ToolResult(success: result.success, data: resultData, error: result.error)
    }

    func resetCosts() {
        totalCost = 0
        callCount = 0
    }

    func getCostSummary() -> (calls: Int, totalCost: Decimal) {
        return (callCount, totalCost)
    }
}

Error Handling

Common Tool Errors

Invalid Parameters

Cause: Parameters don’t match the schema or fail validation.Solution:

// Validate early and provide clear error messages
guard let query = parameters["query"]?.stringValue,
      !query.isEmpty else {
    return ToolResult(
        success: false,
        error: "Parameter 'query' is required and cannot be empty"
    )
}

guard query.count >= 2 && query.count <= 200 else {
    return ToolResult(
        success: false,
        error: "Query must be between 2 and 200 characters"
    )
}

Tool Not Registered

Cause: Agent references a tool that hasn’t been registered with ToolsHandler.Solution:

// Verify registration before orbit creation
let requiredTools = ["custom_tool", "web_search"]
let toolsHandler = ToolsHandler.shared

for toolName in requiredTools {
    let isAvailable = await toolsHandler.isToolAvailable(named: toolName)
    if !isAvailable {
        print("Warning: Tool '\(toolName)' not registered")
        // Register it
        if toolName == "custom_tool" {
            await toolsHandler.registerTool(CustomTool())
        }
    }
}

Execution Timeout

Cause: Tool operation takes too long to complete.Solution:

// Implement timeout handling
let timeout: TimeInterval = 30.0

do {
    let result = try await withThrowingTaskGroup(of: ToolResult.self) { group in
        group.addTask {
            try await self.performOperation(parameters)
        }

        group.addTask {
            try await Task.sleep(nanoseconds: UInt64(timeout * 1_000_000_000))
            throw ToolError.timeout
        }

        let result = try await group.next()!
        group.cancelAll()
        return result
    }
    return result
} catch {
    return ToolResult(
        success: false,
        error: "Operation timed out after \(timeout) seconds"
    )
}

External API Failures

Cause: External service is unavailable or returns an error.Solution:

// Implement retry logic with exponential backoff
func executeWithRetry(
    parameters: Metadata,
    maxAttempts: Int = 3
) async throws -> ToolResult {
    var lastError: Error?

    for attempt in 1...maxAttempts {
        do {
            return try await callExternalAPI(parameters)
        } catch {
            lastError = error

            if attempt < maxAttempts {
                let delay = pow(2.0, Double(attempt - 1))
                try await Task.sleep(nanoseconds: UInt64(delay * 1_000_000_000))
            }
        }
    }

    return ToolResult(
        success: false,
        error: "Failed after \(maxAttempts) attempts: \(lastError?.localizedDescription ?? "Unknown error")"
    )
}

Permission Denied

Cause: Tool lacks necessary permissions (file access, location, etc.).Solution:

override func execute(
    with parameters: Metadata
) async throws -> ToolResult {
    // Check permissions before execution
    guard hasRequiredPermissions() else {
        return ToolResult(
            success: false,
            error: """
            Permission denied: This tool requires location access.
            Please enable location services in System Settings.
            """
        )
    }

    return try await performOperation(parameters)
}

private func hasRequiredPermissions() -> Bool {
    // Check actual permissions
    return CLLocationManager.locationServicesEnabled() &&
           CLLocationManager.authorizationStatus() == .authorizedWhenInUse
}

Error Handling Patterns

final class RobustTool: BaseTool {
    override func execute(
        with parameters: Metadata
    ) async throws -> ToolResult {
        do {
            // Validate parameters
            try validateParameters(parameters)

            // Execute operation
            let result = try await performOperation(parameters)

            // Validate result
            try validateResult(result)

            return ToolResult(success: true, data: result)

        } catch let error as ValidationError {
            // Specific error handling
            return ToolResult(
                success: false,
                error: "Validation failed: \(error.message)"
            )

        } catch let error as NetworkError {
            // Network-specific handling
            return ToolResult(
                success: false,
                error: "Network error: \(error.localizedDescription). Please check your connection."
            )

        } catch {
            // Generic error handling
            return ToolResult(
                success: false,
                error: "Unexpected error: \(error.localizedDescription)"
            )
        }
    }

    private func validateParameters(_ parameters: Metadata) throws {
        // Parameter validation logic
    }

    private func validateResult(_ result: Metadata) throws {
        // Result validation logic
    }
}

Best Practices

Tool Design

Clear Naming

Use descriptive, action-oriented namesGood: send_email, analyze_sentiment, fetch_stock_price Bad: tool1, helper, process

Detailed Descriptions

Write comprehensive descriptions that guide LLM usage

description: """
Search the web for current information.
Use when you need real-time data, news,
or information not in your training data.
Returns up to 10 results with URLs and snippets.
"""

Schema Validation

Define strict schemas with constraints

"limit": JSONSchema(
    type: .integer,
    description: "Results (1-100)",
    minimum: 1,
    maximum: 100
)

Single Responsibility

Each tool should do one thing wellGood: Separate read_file and write_file tools Bad: One file_operations tool that does everything

Idempotency

Design tools to be safely re-executable

// Safe to call multiple times
func createDirectory(path: String) {
    if !FileManager.default.fileExists(atPath: path) {
        try FileManager.default.createDirectory(...)
    }
}

Error Messages

Return actionable error messagesGood: “File not found at ‘/path/to/file.txt’. Check the path and try again.” Bad: “Error 404”

Performance Optimization

Caching
Batching
Parallel Execution
Streaming

Cache frequently accessed data:

final class CachedWebSearchTool: BaseTool {
    private var cache: [String: ToolResult] = [:]
    private let cacheExpiry: TimeInterval = 3600 // 1 hour

    override func execute(
        with parameters: Metadata
    ) async throws -> ToolResult {
        guard let query = parameters["query"]?.stringValue else {
            throw OrbitAIError.invalidToolParameters("Missing query")
        }

        // Check cache
        if let cached = cache[query],
           let timestamp = cached.data?["timestamp"]?.doubleValue,
           Date().timeIntervalSince1970 - timestamp < cacheExpiry {
            return cached
        }

        // Fetch fresh data
        let result = try await performWebSearch(query)

        // Cache result
        var resultWithTimestamp = result
        resultWithTimestamp.data?["timestamp"] = .double(Date().timeIntervalSince1970)
        cache[query] = resultWithTimestamp

        return resultWithTimestamp
    }
}

Security Considerations

Always validate and sanitize tool inputs to prevent security vulnerabilities.

final class SecureDatabaseTool: BaseTool {
    override func execute(
        with parameters: Metadata
    ) async throws -> ToolResult {
        guard let query = parameters["query"]?.stringValue else {
            return ToolResult(success: false, error: "Missing query")
        }

        // 1. Validate query (prevent SQL injection)
        guard isValidQuery(query) else {
            return ToolResult(
                success: false,
                error: "Invalid query: Only SELECT statements are allowed"
            )
        }

        // 2. Use parameterized queries
        let safeQuery = sanitizeQuery(query)

        // 3. Implement rate limiting
        guard await checkRateLimit() else {
            return ToolResult(
                success: false,
                error: "Rate limit exceeded. Try again later."
            )
        }

        // 4. Execute with limited permissions
        let results = try await executeWithReadOnlyAccess(safeQuery)

        // 5. Sanitize output (remove sensitive data)
        let sanitized = sanitizeResults(results)

        return ToolResult(success: true, data: sanitized)
    }

    private func isValidQuery(_ query: String) -> Bool {
        // Only allow SELECT statements
        let trimmed = query.trimmingCharacters(in: .whitespaces).uppercased()
        return trimmed.hasPrefix("SELECT") &&
               !trimmed.contains("DROP") &&
               !trimmed.contains("DELETE") &&
               !trimmed.contains("UPDATE") &&
               !trimmed.contains("INSERT")
    }
}

Troubleshooting

Common Issues

Tool Not Found

Symptom: Agent reports tool is unavailable or not found.Diagnosis:

// Check if tool is registered
let toolsHandler = ToolsHandler.shared
let isRegistered = await toolsHandler.isToolAvailable(named: "my_tool")
print("Tool registered: \(isRegistered)")

// List all registered tools
let allTools = await toolsHandler.registeredToolNames()
print("Available tools: \(allTools)")

Solutions:

Register the tool before creating the orbit
Check for typos in tool name (names are case-sensitive)
Verify tool is imported and compiled
For built-in tools, ensure OrbitAI version is up to date

Parameter Validation Failures

Symptom: Tool execution fails with parameter errors.Diagnosis:

// Log parameters in tool
override func execute(
    with parameters: Metadata
) async throws -> ToolResult {
    print("Received parameters: \(parameters)")

    // Check parameter types
    for (key, value) in parameters {
        print("\(key): \(type(of: value)) = \(value)")
    }

    // Validate against schema
    // ...
}

Solutions:

Ensure parameter names match schema exactly
Check parameter types (string vs int vs array)
Verify required parameters are provided
Add default values for optional parameters
Improve error messages in tool description

Slow Tool Execution

Symptom: Tools take too long to execute.Diagnosis:

// Add timing logs
override func execute(
    with parameters: Metadata
) async throws -> ToolResult {
    let startTime = Date()
    defer {
        let duration = Date().timeIntervalSince(startTime)
        print("Tool execution took \(duration)s")
    }

    return try await performOperation(parameters)
}

Solutions:

Implement caching for repeated queries
Use parallel execution for independent operations
Add timeouts to prevent hanging
Optimize database queries or API calls
Consider batch processing

Tool Returns Incorrect Data

Symptom: Tool executes successfully but returns wrong results.Diagnosis:

// Add detailed logging
override func execute(
    with parameters: Metadata
) async throws -> ToolResult {
    print("Input parameters: \(parameters)")

    let intermediateResult = try await step1(parameters)
    print("After step 1: \(intermediateResult)")

    let finalResult = try await step2(intermediateResult)
    print("Final result: \(finalResult)")

    return ToolResult(success: true, data: finalResult)
}

Solutions:

Verify parameter parsing logic
Check data transformations
Validate external API responses
Add unit tests for tool logic
Review LLM’s parameter generation

Agent Doesn't Use Expected Tool

Symptom: Agent doesn’t select the tool you expect for a task.Diagnosis:

// Check tool availability to agent
print("Agent tools: \(agent.tools)")
print("Task tools: \(task.tools ?? [])")

// Review tool description
let tool = await ToolsHandler.shared.tool(named: "my_tool")
print("Tool description: \(tool?.description ?? "Not found")")

Solutions:

Improve tool description to clarify when to use it
Verify tool is in agent’s or task’s tool list
Add examples in tool description
Adjust task description to hint at tool usage
Check if multiple tools overlap in functionality

Debugging Tools

Create a debugging wrapper for tools:

final class DebugToolWrapper: BaseTool {
    private let wrappedTool: BaseTool
    private let verbose: Bool

    init(wrapping tool: BaseTool, verbose: Bool = true) {
        self.wrappedTool = tool
        self.verbose = verbose
    }

    override var name: String { wrappedTool.name }
    override var description: String { wrappedTool.description }
    override var parametersSchema: JSONSchema { wrappedTool.parametersSchema }

    override func execute(
        with parameters: Metadata
    ) async throws -> ToolResult {
        if verbose {
            print("━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━")
            print("🔧 Tool: \(name)")
            print("📥 Input Parameters:")
            for (key, value) in parameters {
                print("   \(key): \(value)")
            }
            print("⏱️  Started at: \(Date())")
        }

        let startTime = Date()

        do {
            let result = try await wrappedTool.execute(with: parameters)
            let duration = Date().timeIntervalSince(startTime)

            if verbose {
                print("✅ Success: \(result.success)")
                print("⏱️  Duration: \(String(format: "%.3f", duration))s")
                print("📤 Output:")
                if let data = result.data {
                    for (key, value) in data {
                        print("   \(key): \(value)")
                    }
                }
                if let error = result.error {
                    print("❌ Error: \(error)")
                }
                print("━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━")
            }

            return result

        } catch {
            let duration = Date().timeIntervalSince(startTime)

            if verbose {
                print("❌ Exception thrown")
                print("⏱️  Duration: \(String(format: "%.3f", duration))s")
                print("💥 Error: \(error)")
                print("━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━")
            }

            throw error
        }
    }
}

// Usage
let debugTool = DebugToolWrapper(wrapping: MyCustomTool())
await ToolsHandler.shared.registerTool(debugTool)

Next Steps

Explore Built-in Tools

Learn about specific built-in tools and their capabilities

Agent Configuration

Understand how to configure agents with tools

Task Orchestration

Learn how tasks leverage tools for execution

LLM Integration

Understand how LLMs interact with tools

Pro Tip: Start with built-in tools and only create custom tools when you need domain-specific functionality or external integrations not covered by the built-in suite.

Getting started

Core Concepts

Learn

​Overview

Extensible

Type-Safe

Async-Ready

Traceable

Intelligent

Composable

​Key Capabilities

​Tool Architecture

​Built-in Tools

​File & Data Operations

file_reader

file_writer

directory_list

csv_processor

data_analyzer

​Web & Network Tools

web_search

web_scraping

api_caller

geocoding

​Apple Platform Integration

apple_calendar

apple_reminders

local_notifications

core_location

weather_kit

​Computation Tools

calculator

code_executor

chart_generator

​Tool Assignment

​Agent-Level Tools

​Task-Level Tools

​Agent Tools vs Task Tools

​Creating Custom Tools

​Basic Custom Tool

​Advanced Custom Tool Example

​Tool Integration

​With LLM Function Calling

​Multi-Tool Workflows

​Tool Chaining

​Tool Execution

​Execution Lifecycle

​Execution Model

​Timeout Handling

​Advanced Tool Setup

​Tool Registry & Management

​Enabling/Disabling Tools

​Tool Introspection

​Tool Versioning

​Usage and Metrics

​Tool Usage Tracking

​Accessing Tool Metrics

​Performance Monitoring

​Cost Tracking for External APIs

​Error Handling

​Common Tool Errors

​Error Handling Patterns

​Best Practices

​Tool Design

Clear Naming

Detailed Descriptions

Schema Validation

Single Responsibility

Idempotency

Error Messages

​Performance Optimization

​Security Considerations

​Troubleshooting

​Common Issues

​Debugging Tools

​Next Steps

Explore Built-in Tools

Agent Configuration

Task Orchestration

Overview

Key Capabilities

Tool Architecture

Built-in Tools

File & Data Operations

Web & Network Tools

Apple Platform Integration

Computation Tools

Tool Assignment

Agent-Level Tools

Task-Level Tools

Agent Tools vs Task Tools

Creating Custom Tools

Basic Custom Tool

Advanced Custom Tool Example

Tool Integration

With LLM Function Calling

Multi-Tool Workflows

Tool Chaining

Tool Execution

Execution Lifecycle

Execution Model

Timeout Handling

Advanced Tool Setup

Tool Registry & Management

Enabling/Disabling Tools

Tool Introspection

Tool Versioning

Usage and Metrics

Tool Usage Tracking

Accessing Tool Metrics

Performance Monitoring

Cost Tracking for External APIs

Error Handling

Common Tool Errors

Error Handling Patterns

Best Practices

Tool Design

Performance Optimization

Security Considerations

Troubleshooting

Common Issues

Debugging Tools

Next Steps