Agents & Roles

Subagent types for specialized tasks and role-based personas for different workflows.

Agents & Roles

Brainstorm uses a system of specialized subagents and configurable roles to tackle different kinds of work. Subagents are spawned for focused tasks with scoped tool access, while roles shape the assistant's persona and model selection.

Subagent Types

Explore

Investigates codebases, maps architectures, and gathers context. Has read-only tool access -- cannot modify files.

``bash


storm spawn explore "Map the authentication flow from login to token refresh"


Allowed tools: file_read, list_dir, glob, grep, git_log, git_status, web_search
Plan
Creates detailed implementation plans with step-by-step breakdowns, effort estimates, and risk assessment.

`bash


storm spawn plan "Add multi-tenant support to the API layer"


Allowed tools: file_read, list_dir, glob, grep, git_log, plan, web_search
Code
Writes and modifies code. The primary implementation agent with full filesystem write access.

`bash


storm spawn code "Implement the user settings page with form validation"


Allowed tools: All filesystem tools, shell, git tools, web_search
Review
Analyzes code for bugs, security issues, performance problems, and style violations. Read-only with reporting output.

`bash


storm spawn review "Review the PR changes for security issues"


Allowed tools: file_read, glob, grep, git_diff, git_log
General
A flexible agent with no specific specialization. Used for tasks that do not fit neatly into other categories.
Decompose
Breaks complex tasks into smaller, independent subtasks that can be executed by other agents. Used internally by the orchestration pipeline.
Research
Focused on gathering external information -- documentation, APIs, best practices. Heavy use of web tools.
Allowed tools: web_search, web_fetch, file_read, glob, grep
External
Interfaces with external systems through MCP servers and API integrations. Used for tasks involving third-party services.
Roles

Roles configure the assistant's persona, system prompt, and preferred model. Switch roles with the /role command or keyboard shortcut.


Architect
High-level system design, architecture decisions, and technical strategy. Defaults to the most capable available model (Opus 4.6, GPT-5.4).
Senior Developer
Full-stack implementation with deep knowledge of patterns, performance, and testing. Balanced model selection.
Junior Developer
Straightforward implementations with detailed explanations. Uses cost-efficient models for routine coding tasks.
Product Manager
Feature specification, user story writing, and requirements analysis. Focuses on clarity and completeness over technical depth.
QA
Testing strategy, test writing, bug reproduction, and quality assurance. Emphasis on edge cases and coverage.
Custom Agents

Define custom agent profiles in your project's brainstorm.toml:

`toml


[[agents]]
name = "db-specialist"
description = "Database migration and query optimization expert"
system_prompt = "You are a database specialist focused on PostgreSQL..."
allowed_tools = ["file_read", "file_write", "shell", "grep"]
default_model = "claude-opus-4-6"

Agent profiles can also be stored in SQLite for persistence across sessions. The CLI merges TOML and database profiles, with TOML taking precedence on conflicts.