--- title: Subagents sidebar_position: 50 sidebar_label: Subagents --- import Tabs from '@theme/Tabs'; import TabItem from '@theme/TabItem'; Subagents are independent instances that execute tasks while keeping your main conversation clean and focused. Think of them as temporary assistants that handle specific jobs, bringing process isolation and context preservation by offloading work to separate instances.

Subagents Walkthrough

## How to Use Subagents :::tip Autonomous Subagent Creation goose can autonomously decide to use subagents when it determines they would be beneficial for your task - you don't always need to explicitly request them. This happens automatically in autonomous [permission mode](/docs/guides/goose-permissions) (the default). Subagents are disabled in manual approval, smart approval, and chat-only modes. ::: To use subagents, ask goose to delegate tasks using natural language. goose automatically decides when to spawn subagents and handles their lifecycle. You can: 1. **Request specialized help**: "Use a code reviewer to analyze this function for security issues" 2. **Reference specific recipes**: "Use the 'security-auditor' recipe to scan this endpoint" 3. **Run parallel tasks**: "Create three HTML templates simultaneously" 4. **Delegate complex work**: "Research quantum computing developments and summarize findings" 5. **Control extension access**: "Create a subagent with only the developer extension to refactor the code" You can run multiple subagents sequentially or in parallel. | Type | Description | Trigger Keywords | Example | |------|-------------|------------------|---------| | **Sequential** (Default) | Tasks execute one after another | "first...then", "after" | `"First analyze the code, then generate documentation"` | | **Parallel** | Tasks execute simultaneously | "parallel", "simultaneously", "at the same time", "concurrently" | `"Create three HTML templates in parallel"` | ## Monitoring Subagent Activity When goose delegates work to a subagent, you can see the subagent's tool calls in real-time. This transparency helps you understand what the subagent is doing and verify it's on the right track. Subagent tool calls appear as expandable sections within the conversation. Click to expand and see the full details of each tool call, including: - The tool name being invoked - Arguments passed to the tool - Tool output and results Subagent tool calls are displayed inline with visual indicators showing the tool name and extension. For example: ```text [subagent:16] text_editor | developer ``` Each tool call shows: - The subagent identifier (e.g., `subagent:16`) - The tool name (e.g., `text_editor`) - The extension providing the tool (e.g., `developer`) :::info If a subagent fails or times out (5-minute default), you will receive no output from that subagent. For parallel execution, if any subagent fails, you get results only from the successful ones. ::: ## Internal Subagents Internal subagents spawn goose instances to handle tasks using your current session's context and extensions. There are two ways to configure and execute internal subagents: 1. **Direct Prompts** - Quick, one-off tasks using natural language instructions 2. **Recipes** - Reusable, structured configurations for specialized subagent behavior ### Direct Prompts Direct prompts provided for one-off tasks using natural language prompts. The main agent automatically configures the subagent based on your request. **goose Prompt:** ``` "Use 2 subagents to create hello.html with 'Hello World' content and goodbye.html with 'Goodbye World' content in parallel" ``` **Tool Output:** ```json { "execution_summary": { "total_tasks": 2, "successful_tasks": 2, "failed_tasks": 0, "execution_time_seconds": 16.2 }, "task_results": [ { "task_id": "create_hello_html", "status": "success", "result": "Successfully created hello.html with Hello World content" }, { "task_id": "create_goodbye_html", "status": "success", "result": "Successfully created goodbye.html with Goodbye World content" } ] } ``` ### Recipes Use [recipe](/docs/guides/recipes/) files to define specific instructions, extensions, and behavior for subagents. Recipes provide reusable configurations that can be shared and referenced by name. **Creating a Recipe File** `code-reviewer.yaml` ```yaml id: code-reviewer version: 1.0.0 title: "Code Review Assistant" description: "Specialized subagent for code quality and security analysis" instructions: | You are a code review assistant. Analyze code and provide feedback on: - Code quality and readability - Security vulnerabilities - Performance issues - Best practices adherence activities: - Analyze code structure - Check for security issues - Review performance patterns extensions: - type: builtin name: developer display_name: Developer timeout: 300 bundled: true parameters: - key: focus_area input_type: string requirement: optional description: "Specific area to focus on (security, performance, readability, etc.)" default: "general" prompt: | Please review the following code focusing on {{focus_area}} aspects. Provide specific, actionable feedback with examples. ``` **Place your recipe file where goose can find it** - Set [`GOOSE_RECIPE_PATH`](/docs/guides/recipes/recipe-reference#recipe-location) environment variable to your recipe directory - Or place it in your current working directory **goose Prompt** ``` Use the "code-reviewer" recipe to analyze the authentication feature I implemented ``` **goose Output** ``` I'll use your code-reviewer recipe to create a specialized subagent for this analysis. 🤖 Subagent created using code-reviewer recipe 💭 Analyzing authentication function for security issues... 🔧 Scanning code structure and patterns... ⚠️ Security vulnerabilities detected! ## Code Review Results ### Critical Issues Found: 1. **SQL Injection Vulnerability**: Direct string interpolation in SQL query 2. **Missing Password Hashing**: Plain text password comparison ### Recommendations: - Use parameterized queries or ORM - Implement proper password hashing (bcrypt, scrypt) - Add input validation and sanitization ``` ## External Subagents External subagents let you bring in AI agents from other providers and platforms, enabling goose to coordinate and integrate your workflow with the broader ecosystem. In the below example, we use Codex as a subagent by running it as an MCP server: **[goose Configuration File](/docs/guides/config-files)** (`.~/.config/goose/config.yaml `): ```yaml subagent: args: - mcp-server bundled: true cmd: codex description: OpenAI Codex CLI Subagent enabled: true env_keys: - OPENAI_API_KEY envs: {} name: subagent timeout: 300 type: stdio ``` **External Tool Configuration** (`~/.codex/config.toml`): ```toml # Use fast model for quick responses # model = "codex-mini-latest" disable_response_storage = true # Never prompt for approval - auto-execute approval_policy = "never" [sandbox] mode = "workspace-write" ``` **goose Prompt:** ``` "Use the codex subagent to analyze my codebase structure and identify the main components" ``` **goose Output:** ```md Based on my analysis of your codebase, here are the main components: 1. **Core Agent System** (`crates/goose/src/agents/`) - Agent orchestration and session management - Tool execution framework - Extension system integration 2. **CLI Interface** (`crates/goose-cli/`) - Command-line interface and session handling - Configuration management 3. **Server Components** (`crates/goose-server/`) - HTTP API endpoints - WebSocket communication for real-time interaction 4. **Desktop UI** (`ui/desktop/`) - Electron-based desktop application - TypeScript frontend with React components The architecture follows a modular design with clear separation between the core agent logic, interfaces, and UI components. ``` ## Suggested Use Cases **Independent Operations** - Creating multiple files with similar structure - Basic data processing tasks - File transformations and generations **Context Preservation** - Complex analysis that generates lots of tool output - Specialized tasks better handled by dedicated agents - Keeping main conversation focused on high-level decisions **Process Isolation** - Tasks that might fail without affecting main workflow - Operations requiring different configurations - Experimental or exploratory work ## Lifecycle and Cleanup Subagents are temporary instances that exist only for task execution. After the task is completed, no manual intervention is needed for cleanup. ## Configuration Subagents use the following pre-configured settings, but you can override any defaults using natural language in your prompts. ### Default Settings | Parameter | Default | How to Customize | |-----------|---------|------------------| | **Max Turns** | 25 | Use natural language, set `GOOSE_SUBAGENT_MAX_TURNS`, or configure `settings.max_turns` in [recipes](/docs/guides/recipes/recipe-reference#settings) or subagent tool calls | | **Timeout** | 5 minutes | Request longer timeout in your prompt | | **Extensions** | Inherited from parent | Specify which extensions to use in your prompt | | **Return Mode** | All subagent information provided in main session | Specify how much detail you want in your prompt | :::tip Advanced Customization You can also customize subagent behavior by editing the `subagent_system.md` [prompt template](/docs/guides/prompt-templates). ::: ### Customizing Settings in Prompts You can override any default by including the setting in your natural language request: **Examples:** ``` "Use subagents to analyze code, limit each to 5 turns" ``` ``` "Use a research subagent with 30 turns and 20-minute timeout to investigate quantum computing trends" ``` **Environment variable:** Set `GOOSE_SUBAGENT_MAX_TURNS` to change the default max turns for all subagents. ### Extension Control Control which tools and capabilities subagents can access. By default, subagents inherit all extensions from your main session, but you can restrict access for security, focus or performance. **Examples:** ``` "Create a subagent to write a summary, but don't give it file access" ``` ``` "Use a subagent with only code editing tools to refactor main.py" ``` ### Return Mode Control Choose how much information goose provides from its subagents in your main session. **Full Details (Default):** See all tool executions and reasoning steps ``` "Create a subagent to debug this issue - I want to see the full investigation process" ``` **Summary Only:** Get just the final result to keep your conversation clean ``` "Use a subagent to research this topic and summarize the key findings" ``` ## Security Constraints Subagents operate with restricted tool access to ensure safe execution and prevent interference with the main session. ### Allowed Operations Subagents have access to these safe operations: - **Extension discovery**: Search for available extensions to understand what tools are available - **Resource access**: Read and list resources from enabled extensions for context - **Extension tools**: Use tools from extensions specified in recipes or inherited from the parent session ### Restricted Operations The following operations are blocked to ensure subagents remain focused on their assigned tasks without affecting the broader system state: - **Subagent spawning**: Cannot create additional subagents to prevent infinite recursion - **Extension management**: Cannot enable, disable, or modify extensions to avoid conflicts with the main session - **Schedule management**: Cannot create, modify, or delete scheduled tasks to prevent interference with parent workflows :::info Subagents can browse extensions for suggestions but cannot enable them to avoid modifying the parent session. ::: ## Additional Resources import ContentCardCarousel from '@site/src/components/ContentCardCarousel'; import subagentsVsSubrecipes from '@site/blog/2025-09-26-subagents-vs-subrecipes/subrecipes-vs-subagents.png'; import agentCoordination from '@site/blog/2025-08-14-agent-coordination-patterns/agent-coordination.png';