agent
Ask a coding provider to perform a task with explicit prompt, model, effort, and resource choices.
When to use it
Use an agent node for open-ended analysis or implementation. The provider can use its native tools according to its own permissions, so review both the task and the deployment profile before execution.
Example
Rattish: 1
Workflow:
name: Call an agent
Node step:
type: agent
provider: codex
prompt: Explain the project entry points without changing files.
working-dir: .
memory: "none"
finish: passProvider IDs come from the Rattish provider contracts; for example, use claude-code rather than the legacy subscription name claude_code. When model or effort is omitted, compilation resolves the provider contract default. A deployment profile supplies compatible runtime settings without overriding authored provider/model/effort choices.
Use inline prompt or a prompt-path; they are mutually exclusive. Bind values through with and refer to locals in prompt text as . memory selects "none", run, or all; quote "none" so it is a string rather than an absence literal. This field does not turn on Second Brain.
The default successful output is text. Declare output-schema or output-schema-path for typed structured output, and use only one schema source. repair-attempts allows zero to three attempts to repair invalid structured output, which can incur additional provider usage.
Configuration
Configuration fields use hyphens. These are the node-specific fields; shared execution fields apply separately.
| Field | Type | Required / default | Meaning |
|---|---|---|---|
provider | string | Required | Rattish provider ID. |
profile | string or null | Not set | Optional compatible deployment profile. |
model | string | Provider default | Model ID supported by the selected provider. |
effort | string | Provider default | Effort choice supported by that model. |
working-dir | string | "." | Working directory resolved from the workflow project. |
prompt | string | "" | Inline task prompt. |
prompt-path | string or null | Not set | Readable prompt source file. |
skill | string or null | Not set | Selected skill reference. |
memory | none, run, all | "none" | Provider memory scope. |
tools | list of string | [] | Selected provider tools. |
mcp-servers | list of string | [] | Selected MCP server references. |
env | object | {} | Additional string environment values. |
extra-paths | list of string | [] | Additional authored resource paths. |
output-schema | object or null | Not set | Inline successful-output JSON Schema. |
output-schema-path | string or null | Not set | File containing the successful-output schema. |
repair-attempts | integer (min 0, max 3) | 0 | Bounded structured-output repair attempts. |
Inputs and output
Additional local bindings are supported. See Inputs and bindings for delivery and type rules.
Successful output is a string by default, or the value validated by the declared output schema. Select downstream fields only when the schema declares them.
Preflight and limits
The provider must be installed/authenticated as needed, the working directory must exist, and selected prompt/profile resources must be valid. Preflight does not guarantee account allowance or that a provider will complete the task.
Use Raticode rattish check and Raticode rattish preflight before choosing to run the example. The installed contract is available through Raticode rattish docs --format json.
Continue with Node reference, Routes and joins, or Runs and approvals.
