# Oh My OpenCode: Agents & Categories Guide

A comprehensive reference for the 11 AI agents and 8 task categories that power the multi-model orchestration system.

---

## Table of Contents

- [Overview](#overview)
- [Agents](#agents)
- [Primary Agents](#primary-agents)
- [Subagents](#subagents)
- [Categories](#categories)
- [Default Model Configurations](#default-model-configurations)
- [Suggested Workflows](#suggested-workflows)
- [Configuration](#configuration)
- [Best Practices](#best-practices)

---

## Overview

Oh My OpenCode provides **11 specialized AI agents** and **8 task categories** for multi-model orchestration. Each agent is optimized for specific tasks with carefully tuned models, temperatures, and fallback chains.

### Agent Architecture

| Aspect | Description |
|--------|-------------|
| **Primary Agents** | Respect UI model selection (Sisyphus, Atlas, Prometheus) |
| **Subagents** | Use their own fallback chains, ignore UI selection |
| **Factory Pattern** | Each agent exports `createXXXAgent(model: string): AgentConfig` |
| **Metadata** | `XXX_PROMPT_METADATA` with category, cost, triggers |
| **Tool Restrictions** | Fine-grained control over available tools per agent |

---

## Agents

### Primary Agents

Primary agents respect the user's UI-selected model and serve as the main entry points for user interactions.

#### 1. Sisyphus (Main Orchestrator)

Model: anthropic/claude-opus-4-6
Temperature: 0.1
Variant: (follows UI selection)
Mode: primary
Fallback Chain:
  1. kimi-k2.5
  2. glm-4.7
  3. gpt-5.3-codex
  4. gemini-3-pro

**Purpose**: The primary orchestrator agent that coordinates all work. Named after the Greek mythological figure condemned to roll a boulder uphill eternally—Sisyphus never stops working until the task is complete.

**Key Characteristics**:
- Maintains the master TODO list
- Delegates to specialized subagents
- Enforces code quality and project conventions
- Uses 32k thinking budget tokens
- Contextual Grep for exploration (liberally fires `explore` agents)
- LSP-based refactoring

**When to Use**: Every session starts with Sisyphus. It's the default agent for all user interactions.

**Tool Restrictions**: Full access

---

#### 2. Atlas (Master Orchestrator)

Model: anthropic/claude-sonnet-4-5
Temperature: 0.1
Variant: (follows UI selection)
Mode: primary
Fallback Chain:
  1. kimi-k2.5
  2. gpt-5.2

**Purpose**: Alternative orchestrator with different architectural approach. Holds the todo list and manages complex multi-step workflows.

**Key Characteristics**:
- GPT-optimized and Claude-optimized prompt variants
- Model-specific routing based on provider
- Specialized for large-scale project coordination

**When to Use**: Alternative to Sisyphus for users who prefer a different orchestration style.

**Tool Restrictions**: Denied: `task`, `call_omo_agent`

---

#### 3. Prometheus (Planning Agent)

Model: anthropic/claude-opus-4-6
Temperature: 0.1
Variant: (follows UI selection)
Mode: primary
Fallback Chain:
  1. kimi-k2.5
  2. gpt-5.2

**Purpose**: Strategic planning and work plan generation. Operates in Interview/Consultant mode to deeply understand requirements before planning.

**Key Characteristics**:
- **Interview Mode**: Asks clarifying questions before generating plans
- **High Accuracy Mode**: Thorough analysis with explicit uncertainty acknowledgment
- Generates dependency graphs and parallel execution plans
- NEVER writes implementation code—planner only
- Uses plan templates with mandatory sections

**When to Use**:
- Complex multi-step tasks requiring detailed planning
- Before starting large refactors or features
- When requirements are ambiguous

**Tool Restrictions**: Full access (but designed for planning, not implementation)

---

### Subagents

Subagents use their own fallback chains and are invoked by primary agents for specialized tasks.

#### 4. Hephaestus (Autonomous Deep Worker)

Model: openai/gpt-5.3-codex
Temperature: 0.1
Variant: medium
Mode: subagent
Fallback Chain: None (requires gpt-5.3-codex specifically)

**Purpose**: The "Legitimate Craftsman"—autonomous deep worker for goal-oriented execution without step-by-step instructions.

**Key Characteristics**:
- Inspired by AmpCode's deep mode
- Fires 2-5 parallel explore/librarian agents before writing code
- Explores extensively, understands deeply, then acts decisively
- Matches existing codebase patterns
- End-to-end completion with verification evidence
- **No fallback chain**—requires gpt-5.3-codex specifically

**When to Use**:
- Complex problems requiring deep understanding
- Tasks where you want to provide a goal, not a recipe
- Autonomous exploration and implementation

**Delegation Trigger**: Goal-oriented tasks requiring thorough research before action

---

#### 5. Oracle (Strategic Advisor)

Model: openai/gpt-5.2
Temperature: 0.1
Variant: (default)
Mode: subagent
Fallback Chain:
  1. claude-opus-4-6

**Purpose**: Read-only high-IQ consultant for architecture decisions, debugging, and complex problem-solving.

**Key Characteristics**:
- **Read-only consultation**—never writes code
- High-quality reasoning model
- Used for debugging after 2+ failed attempts
- Architecture and design decisions
- Security/performance concerns
- Multi-system tradeoff analysis

**When to Use**:
- Architecture design questions
- Debugging complex issues
- After completing significant work (self-review)
- Security or performance analysis

**Tool Restrictions**: Denied: `write`, `edit`, `task`, `call_omo_agent`

**Delegation Trigger**: "Consult Oracle for [reason]"

---

#### 6. Librarian (Multi-Repo Research)

Model: zai-coding-plan/glm-4.7
Temperature: 0.1
Variant: (default)
Mode: subagent
Fallback Chain:
  1. glm-4.7-free

**Purpose**: Specialized agent for searching remote codebases, retrieving official documentation, and finding implementation examples.

**Key Characteristics**:
- GitHub CLI integration for repository search
- Context7 for official documentation
- Web search for best practices
- Multi-repository analysis
- Searches external resources (not your codebase)

**When to Use**:
- "How do I use [library]?"
- "What's the best practice for [framework feature]?"
- Working with unfamiliar npm/pip/cargo packages
- Finding OSS implementation examples

**Tool Restrictions**: Denied: `write`, `edit`, `task`, `call_omo_agent`

**Delegation Trigger**: External library mentioned → fire librarian background

---

#### 7. Explore (Fast Contextual Grep)

Model: xai/grok-code-fast-1
Temperature: 0.1
Variant: (default)
Mode: subagent
Fallback Chain:
  1. claude-haiku-4-5
  2. gpt-5-mini
  3. gpt-5-nano

**Purpose**: Blazing fast codebase exploration using contextual grep. Searches YOUR codebase, not external resources.

**Key Characteristics**:
- Contextual grep across the filesystem
- Fast and cheap (FREE cost classification)
- Multiple parallel queries
- Finds patterns in your code

**When to Use**:
- "Find all files matching pattern X"
- "Search for implementation of Y"
- Understanding existing codebase structure
- Finding patterns and conventions

**Tool Restrictions**: Denied: `write`, `edit`, `task`, `call_omo_agent`

**Delegation Trigger**: 2+ modules involved → fire explore background

---

#### 8. Multimodal-Looker (Media Analyzer)

Model: google/gemini-3-flash
Temperature: 0.1
Variant: (default)
Mode: subagent
Fallback Chain: None

**Purpose**: Analyzes media files (PDFs, images, diagrams) that require interpretation beyond raw text.

**Key Characteristics**:
- PDF extraction and analysis
- Image interpretation
- Diagram understanding
- Multimodal reasoning

**When to Use**:
- Analyzing screenshots or diagrams
- Extracting information from PDFs
- Visual content that requires interpretation

**Tool Restrictions**: Allowlist: `read` only

---

#### 9. Metis (Pre-Planning Analysis)

Model: anthropic/claude-opus-4-6
Temperature: 0.3
Variant: (default)
Mode: subagent
Fallback Chain:
  1. kimi-k2.5
  2. gpt-5.2

**Purpose**: Pre-planning consultant that analyzes requests to identify hidden intentions, ambiguities, and AI failure points.

**Key Characteristics**:
- Higher temperature (0.3) for creative analysis
- Identifies hidden requirements
- Spots ambiguities early
- Flags potential AI failure points
- Consultant mode for complex tasks

**When to Use**:
- Ambiguous requirements
- Complex task requiring scope clarification
- Before creating work plans

---

#### 10. Momus (Plan Reviewer)

Model: openai/gpt-5.2
Temperature: 0.1
Variant: (default)
Mode: subagent
Fallback Chain:
  1. claude-opus-4-6

**Purpose**: Expert reviewer for evaluating work plans against rigorous clarity, verifiability, and completeness standards.

**Key Characteristics**:
- Reviews Prometheus-generated plans
- Validates clarity and completeness
- Catches gaps and missing context
- Enforces plan quality standards

**When to Use**:
- After Prometheus generates a plan
- Quality assurance before execution
- When a plan seems incomplete

---

#### 11. Sisyphus-Junior (Category-Spawned Executor)

Model: anthropic/claude-sonnet-4-5
Temperature: 0.1
Variant: (default)
Mode: subagent
Fallback Chain: (follows category model)

**Purpose**: Spawned by Sisyphus when delegating to categories. Executes tasks with category-specific model and context.

**Key Characteristics**:
- Category-spawned executor
- Uses category-specified model
- GPT vs Claude prompt variants
- Inherits skills from category configuration

**When to Use**: Automatically spawned by `delegate_task()` with category parameter.

**Tool Restrictions**: Denied: `task`, `call_omo_agent`

---

## Categories

Categories are domain-specific task delegation buckets. Each category maps to a specific model optimized for that domain.

### Category Summary

| Category | Model | Variant | Temperature | Best For |
|----------|-------|---------|-------------|----------|
| `visual-engineering` | google/gemini-3-pro | — | 0.3 | Frontend, UI/UX, design, styling, animation |
| `ultrabrain` | openai/gpt-5.3-codex | xhigh | 0.1 | Genuinely hard, logic-heavy tasks |
| `deep` | openai/gpt-5.3-codex | medium | 0.1 | Goal-oriented autonomous problem-solving |
| `artistry` | google/gemini-3-pro | high | 0.6 | Complex creative problem-solving |
| `quick` | anthropic/claude-haiku-4-5 | — | 0.1 | Trivial tasks, single file changes |
| `unspecified-low` | anthropic/claude-sonnet-4-5 | — | 0.1 | Tasks that don't fit other categories, low effort |
| `unspecified-high` | anthropic/claude-opus-4-6 | max | 0.1 | Tasks that don't fit other categories, high effort |
| `writing` | google/gemini-3-flash | — | 0.2 | Documentation, prose, technical writing |

### Category Details

#### 1. visual-engineering

Model: google/gemini-3-pro
Temperature: 0.3
Variant: (none)
Description: Frontend, UI/UX, design, styling, animation

**Mindset**: Design-first approach
- Bold aesthetic choices over safe defaults
- Unexpected layouts, asymmetry, grid-breaking elements
- Distinctive typography (avoid: Arial, Inter, Roboto, Space Grotesk)
- Cohesive color palettes with sharp accents
- High-impact animations with staggered reveals
- Atmosphere: gradient meshes, noise textures, layered transparencies

**Avoid**: Generic fonts, purple gradients on white, predictable layouts, cookie-cutter patterns.

---

#### 2. ultrabrain

Model: openai/gpt-5.3-codex
Temperature: 0.1
Variant: xhigh
Description: Use ONLY for genuinely hard, logic-heavy tasks

**Mindset**: Strategic advisor
- Bias toward simplicity: least complex solution that fulfills requirements
- Leverage existing code/patterns over new components
- Prioritize developer experience and maintainability
- One clear recommendation with effort estimate
- Signal when advanced approach warranted

**Response Format**:
- Bottom line (2-3 sentences)
- Action plan (numbered steps)
- Risks and mitigations

---

#### 3. deep

Model: openai/gpt-5.3-codex
Temperature: 0.1
Variant: medium
Description: Goal-oriented autonomous problem-solving

**Mindset**: Autonomous executor
- **NOT** an interactive assistant
- Receives a GOAL, not step-by-step instructions
- Figures out HOW to achieve the goal
- Thorough research before any action (5-15 min reading normal)
- Fix hairy problems requiring deep understanding
- Work independently without frequent check-ins

**Critical Rule**: SILENTLY explore codebase extensively BEFORE making ANY changes.

---

#### 4. artistry

Model: google/gemini-3-pro
Temperature: 0.6
Variant: high
Description: Complex problem-solving with unconventional, creative approaches

**Mindset**: Artistic genius
- Push far beyond conventional boundaries
- Explore radical, unconventional directions
- Surprise and delight: unexpected twists, novel combinations
- Rich detail and vivid expression
- Break patterns deliberately when it serves the creative vision

---

#### 5. quick

Model: anthropic/claude-haiku-4-5
Temperature: 0.1
Variant: (none)
Description: Trivial tasks - single file changes, typo fixes, simple modifications

**Mindset**: Efficient execution
- Fast, focused, minimal overhead
- Get to the point immediately
- No over-engineering
- Simple solutions for simple problems

**⚠️ Caller Warning**: This category uses a LESS CAPABLE model. Your prompt MUST be:
- **EXHAUSTIVELY EXPLICIT** - Leave NOTHING to interpretation
- **MUST DO**: List every required action as atomic, numbered steps
- **MUST NOT DO**: Explicitly forbid likely mistakes
- **EXPECTED OUTPUT**: Describe exact success criteria

---

#### 6. unspecified-low

Model: anthropic/claude-sonnet-4-5
Temperature: 0.1
Variant: (none)
Description: Tasks that don't fit other categories, low effort required

**Selection Gate**: Before selecting, verify:
1. Task does NOT fit: quick, visual-engineering, ultrabrain, artistry, writing
2. Task requires more than trivial effort but is NOT system-wide
3. Scope is contained within a few files/modules

**⚠️ Caller Warning**: Provide clear structure with MUST DO, MUST NOT DO, and EXPECTED OUTPUT.

---

#### 7. unspecified-high

Model: anthropic/claude-opus-4-6
Temperature: 0.1
Variant: max
Description: Tasks that don't fit other categories, high effort required

**Selection Gate**: Before selecting, verify:
1. Task does NOT fit: quick, visual-engineering, ultrabrain, artistry, writing
2. Task requires substantial effort across multiple systems/modules
3. Changes have broad impact or require careful coordination
4. NOT just "complex"—must be genuinely unclassifiable AND high-effort

---

#### 8. writing

Model: google/gemini-3-flash
Temperature: 0.2
Variant: (none)
Description: Documentation, prose, technical writing

**Mindset**: Wordsmith
- Clear, flowing prose
- Appropriate tone and voice
- Engaging and readable
- Proper structure and organization

**Use For**: Documentation, READMEs, articles, technical writing

---

## Default Model Configurations

### Complete Agent Configuration Table

| Agent | Primary Model | Temp | Variant | Fallback Chain | Thinking Budget |
|-------|--------------|------|---------|----------------|-----------------|
| **Sisyphus** | anthropic/claude-opus-4-6 | 0.1 | UI-selected | kimi-k2.5 → glm-4.7 → gpt-5.3-codex → gemini-3-pro | 32k |
| **Atlas** | anthropic/claude-sonnet-4-5 | 0.1 | UI-selected | kimi-k2.5 → gpt-5.2 | 32k |
| **Prometheus** | anthropic/claude-opus-4-6 | 0.1 | UI-selected | kimi-k2.5 → gpt-5.2 | — |
| **Hephaestus** | openai/gpt-5.3-codex | 0.1 | medium | None (required) | — |
| **Oracle** | openai/gpt-5.2 | 0.1 | — | claude-opus-4-6 | 32k |
| **Librarian** | zai-coding-plan/glm-4.7 | 0.1 | — | glm-4.7-free | — |
| **Explore** | xai/grok-code-fast-1 | 0.1 | — | claude-haiku-4-5 → gpt-5-mini → gpt-5-nano | — |
| **Multimodal-Looker** | google/gemini-3-flash | 0.1 | — | None | — |
| **Metis** | anthropic/claude-opus-4-6 | 0.3 | — | kimi-k2.5 → gpt-5.2 | — |
| **Momus** | openai/gpt-5.2 | 0.1 | — | claude-opus-4-6 | — |
| **Sisyphus-Junior** | anthropic/claude-sonnet-4-5 | 0.1 | — | Category-based | — |

### Complete Category Configuration Table

| Category | Model | Variant | Temperature | Description |
|----------|-------|---------|-------------|-------------|
| visual-engineering | google/gemini-3-pro | — | 0.3 | Frontend, UI/UX, design |
| ultrabrain | openai/gpt-5.3-codex | xhigh | 0.1 | Hard logic-heavy tasks |
| deep | openai/gpt-5.3-codex | medium | 0.1 | Autonomous problem-solving |
| artistry | google/gemini-3-pro | high | 0.6 | Creative approaches |
| quick | anthropic/claude-haiku-4-5 | — | 0.1 | Trivial tasks |
| unspecified-low | anthropic/claude-sonnet-4-5 | — | 0.1 | Low effort, unclassifiable |
| unspecified-high | anthropic/claude-opus-4-6 | max | 0.1 | High effort, unclassifiable |
| writing | google/gemini-3-flash | — | 0.2 | Documentation, prose |

**Note**: Temperatures are optimized defaults. You can override them via configuration:
- **0.1**: Consistent, deterministic output (logic-heavy tasks)
- **0.2-0.3**: Balanced creativity (writing, visual-engineering)
- **0.6+**: High creativity for artistic tasks
- **Custom**: Any value 0.0-1.0 via `categories.{name}.temperature`

---

## Suggested Workflows

### Workflow 1: Standard Development Task

User Request → Sisyphus
                ↓
    ┌───────────┼───────────┐
    ↓           ↓           ↓
 Explore    Librarian    (if needed)
    ↓           ↓
    └───────────┘
                ↓
    Delegate to Category + Skills
                ↓
    ┌───────────┼───────────┐
    ↓           ↓           ↓
  Coding    Testing     Verification

**Steps**:
1. User makes request → Sisyphus receives it
2. Sisyphus fires parallel background agents:
- `explore` for codebase patterns
- `librarian` for external docs (if library mentioned)
3. Based on findings, Sisyphus delegates to appropriate category
4. Category-spawned agent executes task
5. Verification and evidence collection

---

### Workflow 2: Complex Feature Implementation

User Request → Sisyphus
                ↓
    Consult Prometheus (Planning)
                ↓
    Generate Work Plan with:
    - Dependency Graph
    - Parallel Execution Waves
    - Category + Skills per Task
                ↓
    Review with Momus (optional)
                ↓
    Execute by Waves
    ┌───────────┬───────────┐
    Wave 1     Wave 2     Wave 3
    (parallel) (parallel) (final)

**Steps**:
1. Sisyphus recognizes complexity → Invokes Prometheus
2. Prometheus interviews user for clarity
3. Prometheus generates detailed work plan
4. (Optional) Momus reviews plan quality
5. Sisyphus executes tasks in parallel waves
6. Each task uses appropriate category + skills

---

### Workflow 3: Debugging Complex Issue

Failed Fix #1 → Sisyphus
                    ↓
    Failed Fix #2 (self)
                    ↓
    Consult Oracle ←────── Debug Strategy
                    ↓
    Implement Fix
                    ↓
    Verify & Document

**Steps**:
1. First fix attempt by Sisyphus
2. If fails, second attempt with different approach
3. After 2 failures, consult Oracle
4. Oracle analyzes and recommends strategy
5. Implement Oracle's recommendation
6. Verify fix works, document learnings

---

### Workflow 4: Frontend/UI Task

User Request (UI-related)
                ↓
    Sisyphus detects UI keywords
                ↓
    delegate_task(
      category="visual-engineering",
      load_skills=["frontend-ui-ux", "playwright"]
    )
                ↓
    Gemini 3 Pro executes with:
    - Design-first mindset
    - Bold aesthetic choices
    - Playwright for verification

**Steps**:
1. Sisyphus detects visual/UI intent
2. Delegates to visual-engineering category
3. Loads relevant skills (frontend-ui-ux, playwright)
4. Category agent executes with design-first approach
5. Uses Playwright for visual verification

---

### Workflow 5: Autonomous Deep Work

User Goal (high-level)
                ↓
    delegate_task(
      category="deep",
      prompt="Goal: [objective]"
    )
                ↓
    Hephaestus/GPT-5.3-Codex:
    1. Explores codebase (5-15 min)
    2. Researches external docs
    3. Understands patterns
    4. Implements solution
    5. Verifies completion
                ↓
    Report results with evidence

**Steps**:
1. Provide high-level goal (not step-by-step)
2. Delegate to `deep` category
3. Agent autonomously explores and implements
4. Minimal status updates (trust autonomy)
5. Report completion with summary

---

### Workflow 6: Documentation Writing

User Request (docs)
                ↓
    delegate_task(
      category="writing",
      prompt="Write [type of doc]"
    )
                ↓
    Gemini 3 Flash executes:
    - Clear, flowing prose
    - Appropriate tone
    - Proper structure
                ↓
    Review and refine

---

## Configuration

### Configuring Agents

Add to your `oh-my-opencode.json`:

{
  "agent_overrides": {
    "sisyphus": {
      "temperature": 0.2,
      "thinking": {
        "type": "enabled",
        "budgetTokens": 64000
      }
    },
    "oracle": {
      "model": "anthropic/claude-opus-4-6",
      "temperature": 0.05
    }
  }
}

### Configuring Categories

Categories support extensive configuration:

{
  "categories": {
    "visual-engineering": {
      "model": "google/gemini-3-pro",
      "variant": "high",
      "temperature": 0.15,
      "top_p": 0.9,
      "maxTokens": 8192
    },
    "ultrabrain": {
      "model": "openai/gpt-5.3-codex",
      "variant": "xhigh",
      "reasoningEffort": "high",
      "temperature": 0.05
    },
    "writing": {
      "model": "google/gemini-3-flash",
      "temperature": 0.3,
      "textVerbosity": "high"
    }
  }
}

**Available Category Properties:**

| Property | Type | Description |
|----------|------|-------------|
| `model` | string | Model identifier (e.g., `google/gemini-3-pro`) |
| `variant` | string | Model variant (e.g., `high`, `max`, `xhigh`) |
| `temperature` | number | 0.0-2.0, controls randomness (0.1 for deterministic) |
| `top_p` | number | 0.0-1.0, nucleus sampling parameter |
| `maxTokens` | number | Maximum response tokens |
| `thinking` | object | `{ type: "enabled"\|"disabled", budgetTokens?: number }` |
| `reasoningEffort` | string | `low`\|`medium`\|`high`\|`xhigh` (OpenAI models) |
| `textVerbosity` | string | `low`\|`medium`\|`high` |
| `tools` | object | `{ [toolName]: boolean }` - Enable/disable specific tools |
| `prompt_append` | string | Additional context appended to system prompt |
| `description` | string | Human-readable description shown in task prompts |

### Using Categories in Delegation

// Basic delegation
delegate_task(
  category="visual-engineering",
  prompt="Create a responsive navigation component"
)

// With skills
delegate_task(
  category="ultrabrain",
  load_skills=["git-master"],
  prompt="Refactor the authentication system"
)

// Background execution
delegate_task(
  category="quick",
  run_in_background=true,
  prompt="Fix the typo in README.md"
)

---

## Best Practices

### Agent Selection

| Scenario | Recommended Agent |
|----------|-------------------|
| Starting a session | Sisyphus (default) |
| Complex planning | Prometheus |
| Debugging (2+ failures) | Oracle |
| External library research | Librarian |
| Codebase exploration | Explore |
| PDF/image analysis | Multimodal-Looker |
| Pre-planning analysis | Metis |
| Plan review | Momus |
| Autonomous deep work | Hephaestus (via `deep` category) |

### Category Selection

| Task Type | Recommended Category |
|-----------|---------------------|
| UI/UX work | visual-engineering |
| Hard logic/architecture | ultrabrain |
| Autonomous problem-solving | deep |
| Creative tasks | artistry |
| Single file fixes | quick |
| Documentation | writing |
| Low effort, unclassifiable | unspecified-low |
| High effort, unclassifiable | unspecified-high |

### Temperature Guidelines

| Agent Type | Recommended Temp | Max |
|------------|-----------------|-----|
| Code agents | 0.1 | 0.3 |
| Planning agents | 0.1 | 0.2 |
| Creative agents | 0.2 | 0.5 |
| Analysis agents | 0.1 | 0.3 |

**Anti-pattern**: Never use >0.3 for code agents. It leads to inconsistent output quality.

### Parallel Execution

Always prefer parallel exploration:

// GOOD: Parallel background agents
delegate_task(agent="explore", run_in_background=true, prompt="Find auth patterns")
delegate_task(agent="explore", run_in_background=true, prompt="Find error handling patterns")
delegate_task(agent="librarian", run_in_background=true, prompt="Research JWT best practices")

// Continue work while agents run
// System notifies when each completes

// BAD: Sequential blocking calls
const result1 = await delegate_task(agent="explore", ...)  // Don't do this
const result2 = await delegate_task(agent="librarian", ...) // Don't do this

### Skills Loading

Always evaluate ALL skills for relevance:

// GOOD: Justify inclusions/omissions
delegate_task(
  category="visual-engineering",
  load_skills=["frontend-ui-ux", "playwright"],
  prompt="Build login page"
)

// Skills Evaluation:
// - INCLUDED frontend-ui-ux: Design patterns needed
// - INCLUDED playwright: For verification
// - OMITTED git-master: No commit operations needed

---

## Summary

Oh My OpenCode's agent and category system provides:

1. **11 Specialized Agents** - Each optimized for specific tasks with tuned models and temperatures
2. **8 Task Categories** - Domain-specific delegation with appropriate models
3. **Intelligent Fallbacks** - Graceful degradation when preferred models unavailable
4. **Parallel Execution** - Fire multiple agents simultaneously for efficiency
5. **Flexible Configuration** - Override defaults via JSON configuration
6. **Best Practice Enforcement** - Built-in guardrails for quality output

**Remember**: The key to effective orchestration is understanding when to delegate to which agent/category, and always verifying results rather than trusting "I'm done" reports.

---

*Generated for Oh My OpenCode — Multi-model agent orchestration made simple.*