Documentation Index
Fetch the complete documentation index at: /llms.txt
Use this file to discover all available pages before exploring further.
The ten step types a workflow is built from, what each does, and which ones call a model.
| Step type | Purpose | Calls a model |
|---|---|---|
guided | Reason with a focused instruction and require specified tool calls | Yes |
agent | Open-ended reasoning with a constrained tool set | Yes |
tool | Invoke one declared tool with deterministic arguments | No |
transform | Filter, join, aggregate or reshape data with deterministic SQL | No |
python | Run deterministic Python in the sandbox | No |
conditional | Branch on declared runtime state | No |
wait_for_input | Pause and route on a person’s response | No |
subworkflow | Invoke another saved workflow | Yes or configurable |
map | Fan out work over a runtime collection and join the results | Yes or configurable |
action | Plan and execute a terminal side effect, optionally with approval | Yes or configurable |
guided, agent, subworkflow, map and action steps. Pins on step types that never call a model are rejected. See Models.
| Declaration | Meaning |
|---|---|
| Typed inputs and outputs | Explicit contracts between nodes |
| Required tools | Calls a guided or agent step must make before it can finish |
| Assertions | Row, aggregate, cross-field or artifact checks on results |
| Bounded retries | Maximum attempts when validation fails |
| Fallback instructions | A controlled agentic fallback for recoverable deterministic-step failures |
| Memoization | Reuse of a reproducible step recipe when appropriate |
gate_check, verify, condition and if_false_goto are rejected on write. Current specs carry a compiler version and must pass the same compilation path used at run time.