Agents

Agents

An agent is a ReAct reasoning loop with its own identity, tools, model, and system prompt. Agents think, act, observe, and iterate — handling tasks that require multiple tool calls and real decision-making.

---

Creating an Agent

Go to Settings → Agents and click New Agent. Configure:

Field	Description
Name	Unique identifier shown in the UI and messaging channels
Description	What the agent does — also used for agent selection in multi-agent mode
System Prompt	Defines the agent's persona, expertise, and constraints
Tools	Select which tools the agent can use (all tools, or a restricted subset)
LLM Provider	Which provider to use (Ollama, Anthropic, OpenAI, Gemini, xAI, DeepSeek)
Model	Specific model to use for this agent (overrides global default)
Code Repositories	Link repos for semantic code search and filesystem access

---

Example Agents

Research Agent

{
  "name": "Research Agent",
  "description": "Deep research using web browsing and document parsing",
  "tools": ["browser_navigate", "browser_snapshot", "parse_pdf", "parse_xlsx", "vault_write"],
  "system_prompt": "You are a thorough research analyst. For any research task: browse primary sources, extract key data, parse any documents you find, and save a structured report to the vault."
}

Data Agent

{
  "name": "Data Agent",
  "description": "Analyzes data files and databases, produces reports",
  "tools": ["list_tables", "get_table_schema", "run_sql_query", "execute_python", "vault_write", "vault_read"],
  "system_prompt": "You are a data analyst. Explore the database schema, write SQL queries to extract insights, then use Python (pandas/matplotlib) to analyze and visualize results. Save all outputs to the vault."
}

Developer Agent

{
  "name": "Strict Developer",
  "description": "Writes production-ready code, creates APIs, and runs self-correcting tests",
  "tools": ["execute_python", "mcp_github", "mcp_slack", "vault_write", "vault_read"],
  "system_prompt": "You are a senior backend engineer. Write robust, functional code, execute it using the Python tool to verify logic, and save the final output to the vault."
}

---

Agent Packs

Synapse includes curated collections of pre-built agents. Import them from Settings → Import/Export.

Starter Pack

Get up and running fast:

• Personal Assistant — full tool access, general-purpose
• Web Research Agent — browser + PDF/Excel parsing

Developer Pack

Built for engineering teams:

• Code Review Agent
• Software Engineer Agent
• QA Engineer
• Dev base orchestration

Productivity Pack

Business and content power-users:

• Data Analyst
• Content Writer
• Jira Analyst
• Slack Notifier

MCP Pack

Pre-configured remote MCP server connections.

---

Import / Export

Export agents (and their associated orchestrations, MCP server configs) as a portable .bundle.json file via Settings → Import/Export. Import bundles from other Synapse instances or share them with your team.

Preview before commit — Synapse shows you exactly what will be imported before you confirm.

---

Multi-Agent Mode in Messaging

When connected to a messaging platform, users can switch the active agent mid-conversation:

/agent Research Agent    # switch to a specific agent
/agents                  # list all available agents

See Messaging Integration for setup details.

---

Tips

• Restrict tools for specialized agents — a SQL agent doesn't need browser access
• Use model overrides to run expensive models only where needed (e.g., analysis steps), and cheap models for routing
• System prompts matter — be specific about output format, what to save to vault, and when to ask for clarification
• Link code repos to agents that do code review or development — enables semantic code search