Understanding Agents

Agents are specialized AI assistants you configure for specific tasks. Think of them as team members with different expertise: one for research, another for code review, another for writing documentation.

What Makes an Agent?

Each agent has:

Component	Purpose
System Prompt	Instructions that define its personality and expertise
Model	Which AI model powers it (Claude Sonnet 5.0, Opus 4.6, Gemini 3 Pro, GPT-5.3, or local models)
Adapter	Which AI backend to use (claude-code, google-gemini, openai-codex, opencode, lmstudio, ollama)
Tools	What actions it can perform (read files, search web, run commands)
Permissions	Which actions need your approval before executing
MCP Servers	Which MCP servers this agent can access (e.g., sciorex-tickets, sciorex-research)
Thinking Level	Extended thinking depth (Off, Think, Think Hard, Think Harder, Ultrathink — Claude only)
Effort Level	How much reasoning to apply (Low, Medium, High, Max — Claude Opus 4.6 only)
Input/Output Schema	Structured JSON schemas defining expected input and output formats
Labels	Tags for organizing and filtering agents

When to Use Agents vs Plain Chat

Use Agents when...	Use Plain Chat when...
You have a recurring task type	You have a one-off question
You want consistent behavior	You want flexibility
You're building automated flows	You're exploring ideas
Multiple people need the same assistant	You're just you
You need multi-model debate	You're working with one model

Creating Your First Agent

1. Go to Agents in the sidebar
2. Click New Agent
3. Give it a name and description
4. Write a system prompt explaining what it should do
5. Select which tools it can use
6. Save

Example prompt for a research agent:

You are a research assistant. When given a topic:
1. Search for recent papers and articles
2. Summarize the key findings
3. Note any controversies or open questions
4. Suggest further reading

Always cite your sources.

Tool Permissions

Agents can use various tools, but you control which ones:

Safe to auto-approve:

• Read (viewing files)
• Glob (finding files)
• Grep (searching in files)
• WebSearch (searching online)

Require approval:

• Write (creating/modifying files)
• Edit (changing existing files)
• Bash (running shell commands)

You can configure permissions at multiple levels:

• Auto-approve all (autoApprove: true) — all tool calls run without asking
• Auto-approve specific tools (autoApproveTools) — only listed tools run automatically
• Require human approval (requiresHumanApproval) — listed tools always need your confirmation

WARNING

Be cautious with auto-approving Bash commands. An agent with unrestricted shell access can modify your system.

Community Agents

Sciorex provides access to a community marketplace with pre-configured agents you can install:

Agent	What it does
Orchestrator	Breaks down complex tasks and coordinates work
Planner	Creates step-by-step implementation plans
Executor	Implements code changes
Tester	Writes and runs tests
Documenter	Creates documentation
Reviewer	Reviews code for issues

You cannot edit community agents directly, but you can duplicate them as a starting point for your own.

Council Mode

Council Mode lets multiple AI models debate a topic in structured rounds. Instead of one agent, you get perspectives from several models simultaneously.

• Configure which models participate
• Set the number of debate rounds
• Use custom personas with YAML import/export
• Review the synthesized conclusion

Council Mode is ideal for architecture decisions, code review debates, and research analysis where multiple viewpoints add value.

See Council Mode for full documentation.

Agent Teams

Agent Teams allow multiple agents to collaborate on complex tasks:

• Assign roles (lead, researcher, implementer, reviewer) with color-coded identification
• Agents communicate via a mailbox system and hand off work
• Teammates can operate in plan mode (requiring approval before implementation)
• Track team progress with subagent history, duration, and token usage
• Monitor spawned subagents with real-time status (running, stopping, completed, failed, cancelled, killed)

This is more powerful than sequential agent runs — agents actively coordinate and share context.

Sessions and History

Every conversation with an agent creates a session. Sessions:

• Save the full conversation history
• Can be linked to tickets for tracking
• Can be forked to explore different approaches (using --resume and --fork-session)
• Support labels for organization
• Track permission modes per session (Normal, Auto-approve edits, YOLO)
• Are stored locally in your workspace

Next Steps

• Your First Agent - Step-by-step tutorial
• Code Review Agent - Practical example
• Full agent configuration reference
• Models comparison - Choose the right model
• Using agents in automated flows
• Council Mode - Multi-model structured debate
• Agent Teams - Collaborative multi-agent work