Built-in Agents
QARK ships with a default agent and a set of built-in agents organized into two tiers: orchestrator agents that you invoke directly, and tool agents that orchestrators dispatch as sub-agents.
The default agent
Section titled “The default agent”Every conversation that has no explicit agent assigned uses the default agent. This is not a built-in agent in the gallery — it is a hardcoded system prompt that provides baseline behavior:
- Responds with markdown formatting (headings, lists, tables, code blocks with language identifiers, LaTeX math)
- Direct and concise communication — leads with the answer, matches depth to question complexity
- Technical accuracy: working code examples (not pseudocode), version numbers, specific function names
- Problem solving: numbered steps for complex problems, starts debugging with the most likely cause
- Conversation-aware: references earlier messages, adapts when the topic changes
The default agent inherits whatever model, tools, and context strategy the conversation is configured with. To override its system prompt, either assign a custom agent or add a conversation-level system prompt override in the Config tab (append or replace mode).
Built-in agent rules
Section titled “Built-in agent rules”Built-in agents carry an is_builtin flag:
- Read-only — the system prompt, tool selection, model assignment, and parameters are locked
- Persistent — built-in agents survive updates. They are never removed by an upgrade, but their definitions update automatically if the shipped configuration changes
- Duplicatable — select any built-in agent and choose Duplicate to create an editable copy with all the original’s settings
To customize a built-in agent, duplicate it first. Your copy is fully editable — change the system prompt, swap the model, add or remove tools, rename it. The original remains untouched.
Orchestrator agents
Section titled “Orchestrator agents”These are full agents you assign to conversations. Each dispatches specialist tool agents to complete multi-step workflows.
Deep Researcher
Section titled “Deep Researcher”| Icon / Color | 🔬 #dc2626 (red) |
| Tools | thinking, web search, web fetch |
| Sub-agents | Query Decomposer, Web Researcher, Source Deep Diver, Fact Verifier, Comparative Analyst, Report Synthesizer |
Breaks a research question into sub-queries, dispatches specialists to search the web, deep-dive into sources, verify facts, run comparative analysis, and synthesize a final report.
Research Director
Section titled “Research Director”| Icon / Color | 💼 #8b5cf6 (purple) |
| Tools | web search, document search, image generation |
| Sub-agents | Research Analyst |
Coordinates research with document analysis. The Research Analyst sub-agent has its own sub-agent (Fact Checker), creating a two-level delegation chain.
Codebase Paper Writer
Section titled “Codebase Paper Writer”| Icon / Color | 📖 #0891b2 (cyan) |
| Tools | thinking, combine files, save paper PDF |
| Sub-agents | Codebase Analyzer, Web Researcher, Academic Researcher, Paper Outliner, Section Writer, Paper Editor, Citation Formatter |
Analyzes a codebase, researches related academic work, and produces a structured paper with sections, citations, and PDF output.
Review Article Writer
Section titled “Review Article Writer”| Icon / Color | 📑 #7c3aed (violet) |
| Tools | thinking, combine files, save paper PDF |
| Sub-agents | Web Researcher, Academic Researcher, Source Deep Diver, Paper Outliner, Section Writer, Paper Editor, Citation Formatter |
Researches a topic across web and academic sources, then produces a review article with proper structure, citations, and PDF export.
Tool agents
Section titled “Tool agents”These agents have agent_type: "tool" — they are designed to be called by orchestrators or other agents, not invoked directly. Each has typed input/output schemas and focused tool access. See Agent-as-Tool for how tool agents work.
Research and analysis
Section titled “Research and analysis”| Agent | Icon | Color | Tools | Purpose |
|---|---|---|---|---|
| Query Decomposer | 🧩 | #6366f1 | thinking | Breaks complex questions into targeted sub-queries |
| Web Researcher | 🔍 | #0ea5e9 | web search, web fetch, thinking | Searches the web and synthesizes findings |
| Academic Researcher | 🎓 | #7c3aed | web search, web fetch, thinking | Searches for academic papers and scholarly sources |
| Source Deep Diver | 🤿 | #14b8a6 | web fetch, thinking | Extracts detailed information from specific URLs |
| Fact Verifier | ✅ | #22c55e | web search, web fetch, thinking | Cross-references claims against multiple sources |
| Comparative Analyst | ⚖️ | #f59e0b | web search, web fetch, thinking | Compares alternatives across defined criteria |
| Report Synthesizer | 📝 | #8b5cf6 | thinking | Combines findings from multiple agents into a coherent report |
| Fact Checker | 🛡️ | #10b981 | web search, web fetch | Verifies factual claims (used as sub-agent by Research Analyst) |
| Research Analyst | 🕵️ | #3b82f6 | web search, web fetch, document search, + Fact Checker | Conducts research with built-in fact-checking via its own sub-agent |
| Codebase Analyzer | 🔎 | #06b6d4 | unix commands, document search, thinking | Analyzes code structure and patterns via filesystem tools |
Academic writing
Section titled “Academic writing”| Agent | Icon | Color | Tools | Purpose |
|---|---|---|---|---|
| Paper Outliner | 📋 | #8b5cf6 | thinking, unix commands, document search | Generates structured paper outlines from research material |
| Section Writer | ✍️ | #f59e0b | thinking, unix commands, document search | Writes individual paper sections from outlines |
| Paper Editor | 📝 | #10b981 | thinking, unix commands | Edits and refines paper drafts for coherence and style |
| Citation Formatter | 📚 | #6366f1 | thinking, unix commands | Formats references and citations in proper academic style |
Agent references
Section titled “Agent references”Tool agents reference each other using the format agent_tool:@Agent Name. During seeding, these names are resolved to UUIDs. This means:
- The Research Analyst has Fact Checker as a sub-agent — a built-in two-level chain
- The Deep Researcher orchestrates 6 tool agents in a single workflow
- The Codebase Paper Writer coordinates 7 tool agents across research, analysis, and writing phases
Duplicate and customize
Section titled “Duplicate and customize”To create your own variant:
- Open the agent gallery
- Select the built-in agent
- Choose Duplicate — a new editable agent appears with “(Copy)” suffix
- Modify any field — system prompt, tools, model, parameters, icon, color
- Rename and save
| Starting from | Customization | Result |
|---|---|---|
| Deep Researcher | Remove comparative analysis, add document search | Focused Document Researcher |
| Codebase Paper Writer | Swap Academic Researcher for custom internal-docs agent | Internal Architecture Documenter |
| Web Researcher | Add MCP tools for internal databases, restrict to company domains | Internal Knowledge Researcher |
| Report Synthesizer | Change output format to executive summary style | Executive Briefing Agent |