Agent Templates

Agent Templates provide a reusable blueprint system for creating AI agents. Browse pre-built templates, customize them with your own instructions and integrations, or build entirely new ones.

How It Works

1. Templates are blueprints — they define the structure and capabilities
2. Instances are working copies — customized for your specific use case
3. Memory persists across sessions — agents learn and improve over time

Browsing Templates

Navigate to Agent Templates in the sidebar to access the gallery.

Template Categories

Templates are organized into 15 categories:

Template Scopes

Creating an Instance

Instance Configuration

Execution Modes

For goal-oriented mode, provide:

• Goal — What the agent should accomplish
• Success criteria — How to determine completion (LLM judge, output matching, tool success)
• Max iterations — Safety limit (default: 10, max: 50)

Data Sources

Connect external knowledge sources for RAG:

• Google Drive — Documents and spreadsheets
• Notion — Pages and databases
• Confluence — Documentation spaces
• GitHub — Repository contents
• Slack — Channel history
• Microsoft Graph — Microsoft 365 documents

OAuth-based sources require a one-time "Connect" step. Integration-based sources use API keys.

Tools

Enable capabilities for your agent:

Built-in tools:

• Web search
• Code interpreter
• Image generation

MCP tools: Any tools from your configured MCP servers can be attached. Tools are referenced by their MCP config ID.

Triggers

Define how the agent is activated:

Version History

Instances support versioning for safe iteration:

• Edits create new versions — Significant changes create a new version automatically
• Old versions are archived — Previous versions are preserved
• Rollback anytime — Restore any archived version (creates a new version)
• Memory persists — Learning carries forward across versions
• Same stable ID — All versions of an instance share the same tracking ID

Skills

Skills are modular capabilities that can be attached to templates and instances.

What is a Skill?

A skill is a self-contained set of instructions in Markdown format. When attached to an agent, the skill's content becomes part of the agent's memory as a file (skills/{slug}/SKILL.md).

Attaching Skills

1. Open a template's detail page
2. Go to the Skills section
3. Browse available skills and click Attach
4. Set the sort order and whether the skill is required

Skills follow scope rules:

• System skills can attach to any template
• Organization skills only attach to organization templates in the same org
• User skills only attach to your personal templates

Skill Materialization

When you create an instance from a template with skills:

1. Each skill is copied into the instance's memory as a file
2. The agent can read and reference these skill files
3. You can edit materialized skills within the instance without affecting the original

Memory System

Each agent instance has a persistent memory system that stores files the agent can read, write, and learn from.

Memory File Types

Browsing Memory

The Memory tab shows a file tree browser:

• Navigate folders and files
• Click a file to view its content in a syntax-highlighted editor
• Edit files directly to customize agent behavior

Pending Edits

Agents can propose changes to their own memory (e.g., learning from interactions):

1. The agent proposes a file creation, update, or deletion
2. The edit appears in the Pending tab with a reason
3. Review the proposed changes with a diff viewer
4. Approve to apply or Reject to discard

Auto-Approve Mode

Enable auto-approve (YOLO mode) on an instance to let the agent modify its own memory without human review. Useful for agents that need to learn rapidly.

Memory Operations

• Read any memory file
• Write new files or edit existing ones
• Delete files no longer needed
• Initialize memory from template (if starting fresh)
• Export all memory as JSON for backup

Mentioning Agents

In chat interfaces and documents, type @ to mention an agent instance. The search finds instances by:

• Instance name and description
• Attached skills
• Enabled integrations and tools

Best Practices

Template Design

• Be specific — Clear instructions produce better results
• Use sections — Role, domain knowledge, process, examples
• Define use cases — Help users understand when to use the template
• Add pro tips — Guide users on customization

Instance Management

• Start with defaults — Test before heavily customizing
• Review memory edits — Approve useful learning, reject mistakes
• Use versions — Create new versions for major changes
• Connect relevant data — More context means better results

Skills

• Keep skills focused — One capability per skill
• Use descriptive slugs — research-methodology not skill-1
• Test independently — Verify a skill works before attaching broadly

---

Next Steps