Presets
Agent presets as shared capability bundles for Action Agents and ActionFlow AI nodes, including what they copy and what stays on the agent.
Presets
An agent preset is a named, versioned bundle of capabilities. It is not a copy of an agent. The agent still owns its model, system prompt, triggers, and memory.
Presets are workspace records. REST, the TypeScript SDK, and hosted MCP do not create or attach them. Public agent metadata does not include the preset.
Scopes
| Scope | Behavior |
|---|---|
organization | Belongs to one organization. Owner and Admin can create, read, update, and delete. Members can create, read, and update. |
platform | organizationId is null. The editor treats the preset as read-only. |
Cloning a preset creates a new organization preset. The clone copies Knowledge collections, prompt skills, platform skills, tools, and MCP definitions together with the organization's MCP adoption. The clone does not copy database grants, the lead reviewer, or guardrail overrides.
Soft delete is blocked while an agent or a Flow node is still attached. A slug can be reused after soft delete.
What a preset carries
The editor sections are general, knowledge, skills, MCP, database, tools, reviewer, and guardrails.
Shared across agents and supported ActionFlow AI nodes:
- Knowledge collections
- Organization and platform skills
- Named HTTP tools and builtin tool selection
- MCP connections and enabled tool names
Agent-only:
- Database grants. Execution checks the stored introspection fingerprint again.
- Lead reviewer. An organization preset lead replaces the agent's own lead columns. A platform preset lead is not resolved, so the agent keeps its own reviewer. See Lead Review.
Guardrail overrides on a preset are organization policy for agent_preset. They can only be stricter than the floor. They do not increment configVersion. See Guardrails.
Credential secrets are not part of the resolved configuration.
Attachment
An agent has at most one preset. An ActionFlow node instance has at most one preset.
The effective configuration merges preset resources with the target's own resources. On an id collision the preset row wins. A resource is required if either side marks it required. A missing required resource fails resolution. A missing optional resource is dropped with a warning. Local additions are optional.
ActionFlow AI nodes drop database grants and the reviewer as unsupported. Which shared capabilities a node can use depends on the node:
| Node kind | Preset capabilities |
|---|---|
| Generate Text, Generate JSON, Generate File | Knowledge, skills, tools, and MCP |
| Image, speech, video, and transcribe | Knowledge |
| Sentiment and classification | Knowledge and skills |
An Action Agent node is not this filter. Database grants and the lead reviewer apply on the agent path, including the pause described in Lead Review.
The MCP runtime fingerprint includes presetConfigVersion and the resolved Knowledge, skills, tools, and MCP resources. Guardrail policy version stays outside that fingerprint. See MCP.