Core Concepts

Understanding the fundamental building blocks of Agent Workshop.

Domains

A domain represents the area of expertise for your agent. Agent Workshop supports five domains:

DomainDescriptionExample Use Cases
DevelopmentSoftware engineering tasksCode review, testing, debugging, modernization
BusinessBusiness process automationDocument processing, reports, data entry
CreativeContent creationBlog writing, social media, copywriting
DataData analysis and MLVisualization, analysis, ML pipelines
KnowledgeResearch and synthesisLiterature review, citations, research

Your domain choice influences which templates are shown and which tools are recommended.

Templates

A template is a pre-configured starting point for your agent. Templates include:

  • Default tool selections appropriate for the use case
  • Sample prompts demonstrating the agent's capabilities
  • Custom instructions that shape the agent's behavior

Agent Workshop includes 17 templates across all domains. See the Templates Reference for the complete list.

Tools

Tools are the capabilities your agent can use to interact with the world. Each tool has a risk level:

Risk LevelDescriptionExamples
LowRead-only operationsread-file, find-files, search-files, web-search
MediumWrite operations with safeguardswrite-file, edit-file, git-operations, database-query
HighUnrestricted system accessrun-command

Security consideration

High-risk tools like run-command can execute arbitrary shell commands. Only enable these when necessary and understand the implications.

Permission Levels

Permission levels control which tools are enabled by default:

  • Restrictive - Only read-only tools (read-file, find-files, search-files, web-search)
  • Balanced - Read operations plus controlled writes (adds write-file, git-operations, web-fetch, api-client)
  • Permissive - All tools enabled including run-command

MCP Servers

Model Context Protocol (MCP) servers extend your agent's capabilities by connecting to external services. MCP servers can provide:

  • Additional tools (e.g., GitHub operations, database queries)
  • Data sources (e.g., Google Drive, AWS)
  • Custom integrations

Agent Workshop supports four transport types for MCP:

  • stdio - Local command execution (most common)
  • http - REST endpoints
  • sse - Server-Sent Events
  • sdk - In-process JavaScript modules

The 5 Levers (Claude Code)

Claude Code provides five control mechanisms called "levers" that shape agent behavior. These are configured in the target project where you run your agent, not bundled with the generated agent itself.

LeverFile LocationPurpose
MemoryCLAUDE.mdPersistent context loaded every conversation
Slash Commands.claude/commands/*.mdUser-invoked prompt templates
Skills.claude/skills/*/SKILL.mdModel-invoked capabilities
Subagents.claude/agents/*.mdSpecialized delegate agents
Hooks.claude/settings.jsonEvent-driven automation

Levers are project-specific

These files are created in the project where you use Claude Code (or your generated agent), not in the agent itself. This allows you to customize behavior per-project.

See The 5 Levers for detailed documentation on each lever.

Generated Project

When you complete the wizard, Agent Workshop generates a complete TypeScript project that includes:

  • CLI entry point - Interactive command-line interface
  • Agent logic - Core agent implementation with tool bindings
  • Configuration - Settings management and environment variables
  • Permissions - Security boundary enforcement
  • MCP integration - External server connections
  • Workflow commands - Domain-specific operations

The generated code is fully customizable. See Project Structure for details on each file.

SDK Providers

Agent Workshop supports two AI providers:

Claude Agent SDK

  • Built by Anthropic specifically for agent development
  • Native streaming support
  • Built-in file operation tools
  • Models: Claude Sonnet 4.5, Haiku 4.5, Opus 4.1

OpenAI Agents SDK

  • Built by OpenAI for agent development
  • Function calling with tool use
  • Streaming support
  • Models: GPT-5.1, GPT-5 mini, GPT-4.1, and more

Both SDKs produce functionally similar agents. Choose based on your preferred AI provider and model availability.