Call Child Workflows

Updated 3 days ago • October 4, 2026

Use useWorkflow(...) when a node needs to run another registered workflow, receive its result, and continue. The call belongs in node code or a custom hook. The parent graph does not need a special node type or a declared call edge.

transitionTo hands off execution to another root workflow. It does not provide a return value to the calling node. Use useWorkflow for a call that returns.

Define the child's input and result

Callable workflows declare inputSchema and outputSchema. These describe the workflow's data contract, not which parents may call it. Both parent and child must be registered with the agent.

import { createAgent, createWorkflowHooks, defineWorkflow } from "kortyx";
import { z } from "zod";

const ResearchInput = z.object({ topic: z.string() });
const ResearchOutput = z.object({
  summary: z.string(),
  sources: z.array(z.string()),
});

export const researchWorkflow = defineWorkflow({
  id: "research-workflow",
  version: "1.0.0",
  inputSchema: ResearchInput,
  outputSchema: ResearchOutput,
  nodes: {
    summarize: {
      run: async ({ input }: { input: z.output<typeof ResearchInput> }) => {
        // Replace this illustrative result with your research service/model call.
        return {
          data: { summary: `Research about ${input.topic}`, sources: [] },
        };
      },
    },
  },
  edges: [
    ["__start__", "summarize"],
    ["summarize", "__end__"],
  ],
});

const { useWorkflow } = createWorkflowHooks({
  "research-workflow": researchWorkflow,
});

export async function useResearch(args: { id: string; topic: string }) {
  const result = await useWorkflow({
    id: args.id,
    workflow: "research-workflow",
    input: { topic: args.topic },
  });
  return result.data;
}

export const reportWorkflow = defineWorkflow({
  id: "report-workflow",
  version: "1.0.0",
  nodes: {
    report: {
      run: async ({ input }: { input: unknown }) => {
        const research = await useResearch({
          id: "research",
          topic: String(input),
        });
        // This code runs after the child has completed.
        return { data: { report: research.summary.toUpperCase() } };
      },
    },
    finish: {
      run: () => ({ data: { reportReady: true } }),
    },
  },
  edges: [
    ["__start__", "report"],
    ["report", "finish"],
    ["finish", "__end__"],
  ],
});

export const agent = createAgent({
  workflows: [reportWorkflow, researchWorkflow],
  defaultWorkflowId: "report-workflow",
});

The child's last node returns an ordinary { data: ... } and reaches __end__. There is no return node or returnTo command. On completion, Kortyx validates the child's accumulated GraphState.data using outputSchema; that parsed value becomes result.data.

An ordinary Zod object exposes its declared fields. Use .passthrough() when additional accumulated data should be returned. Parent data is updated only through the caller's own node result; the child does not automatically merge its state into its parent.

Type safety

The bound hook infers input from the selected workflow's input schema and result.data from its output schema. Invalid workflow IDs, wrong inputs, and unknown output fields are TypeScript errors. Use the same definition objects for the hook registry and the agent registry.

You can also pass a typed definition directly without binding string IDs:

The unbound hook accepts dynamic string IDs, but their result is Record<string, unknown>. Use a bound registry or a typed reference for inferred properties. Do not add a caller-supplied result cast to simulate type safety.

Runtime validation remains necessary: a node implementation or external service can produce invalid data even when the caller is typed. Input transforms run on the initial invocation; resume uses the parsed input saved in the child checkpoint. Output transforms run at completion, and cached replay returns that parsed result.

Keep definitions inferred from defineWorkflow(...). Widening them to WorkflowDefinition loses the concrete schema information needed for typed calls. Schemas are TypeScript/JavaScript objects; plain YAML/JSON workflow files cannot embed executable Zod schemas.

Human interrupts and resume

Inside a child, use the same useInterrupt(...) and useReason({ interrupts: { contracts } }) APIs used in a root workflow. Custom contract requests and structured responses bridge through the parent. The singular interrupt option is deprecated and removed next major. For example, this can be the research workflow's final node:

import { useInterrupt } from "kortyx";

export async function reviewResearchNode({
  input,
}: {
  input: { summary: string; sources: string[] };
}) {
  const approved = await useInterrupt({
    id: "review-research",
    request: {
      kind: "choice",
      question: "Use this research?",
      options: [
        { id: "yes", label: "Use research" },
        { id: "no", label: "Discard research" },
      ],
    },
  });
  return {
    data: {
      summary: approved === "yes" ? input.summary : "Research declined",
      sources: approved === "yes" ? input.sources : [],
    },
  };
}

The parent remains suspended at await useWorkflow(...). The application receives the normal interrupt stream chunk and resumes using its token and request ID. There is no separate child-resume endpoint. Existing React interrupt rendering continues to work.

Child graph checkpoints travel inside the parent's saved hook state. When the parent node replays, the hook resumes the waiting child or returns a previously completed child's cached result. Completed child nodes do not repeat just because the parent resumes. Previous bridged interrupts retain their positions even if later children or the parent also interrupt.

Code surrounding a hook can replay, including code in catch and finally. Keep call order and input stable. Use idempotency keys for external writes; a flag in runtime state alone does not make a write atomic with a checkpoint. A child failure rejects with WorkflowCallError; a parent can handle that error, but catching a suspension cannot turn a waiting call into a successful fallback.

Run independent children in parallel

Use parallel from kortyx inside a node. Each child keeps its own schemas and stable call ID; the returned tuple follows the input order:

Register both schema-bearing children with the parent agent. Await each group before starting a dependent group in the same node. A child can itself call parallel for its own children. Keep group order, membership, call IDs and inputs stable through replay. Call the helper directly around the child calls so it owns them before they dispatch.

The join waits for every sibling to finish, fail or suspend before exposing a parent pause. If several children interrupt, all their snapshots are saved; their questions are presented one at a time through the existing parent resume handle. Completed and failed children are cached, and each answer is routed to its own child. A slow running sibling delays the parent suspension. Graph and external-effect durability remain checkpoint-based.

For example, if company research asks for approval at 2 seconds and role research finishes at 60 seconds, the parent exposes the approval at about 60 seconds. Approvals are not delivered immediately while siblings continue running, and a suspended result does not expose a batch of child questions. Answer the current question through the parent, then use the fresh handle returned with the next question. Concurrent child execution does not imply simultaneous approval delivery.

If children fail, parallel throws ParallelError after the group settles. Its errors contains the failures and its results contains standard fulfilled/rejected entries in input order, allowing application reconciliation. A waiting sibling is preserved before terminal failures are delivered. Suspension, cancellation and execution-limit exhaustion propagate as control flow and cannot become successful fallback output. Use instanceof ParallelError when catching task failures; do not swallow other errors.

Children share the root signal and node/model/tool/child allowances. Cached work is not charged again. Continue remains an explicit server-authorized allowance decision. This helper has no concurrency-cap option: its array contains eager calls. Overlapping groups in the same node and native Promise.all child calls remain unsupported.

To preserve dependency waves, await one parallel group before starting the next inside the parent node. Independent tasks in a wave still run concurrently; this pattern does not require parallel graph edges.

Child calls in parallel graph branches

Forward-moving parallel graphs can call children from their ordinary nodes. Callable children can also contain parallel graph edges. Keep using useWorkflow; no wrapper workflow, call edge, or new resume endpoint is required.

nodes: {
  company: {
    run: async () => ({
      data: { company: (await useWorkflow({
        id: "research", workflow: companyResearchWorkflow, input: { companyId },
      })).data },
    }),
  },
  role: {
    run: async () => ({
      data: { role: (await useWorkflow({
        id: "research", workflow: roleAnalysisWorkflow, input: { roleId },
      })).data },
    }),
  },
  plan: { run: planNode },
},
edges: [
  ["__start__", "company"], ["__start__", "role"],
  ["company", "plan"], ["role", "plan"], ["plan", "__end__"],
],

The two calls run independently. Call IDs need only be unique within their own node activation. plan waits for both selected predecessors, even when one branch has more nodes. Unselected conditional paths do not count as required predecessors. Nodes receive their dependency results, not unrelated sibling outputs.

Disjoint top-level fields combine automatically. If unrelated branches both write the same field, the shared node fails before its code runs with GRAPH_OUTPUT_CONFLICT, naming the field and writers. The runtime never renames fields or picks a winner. Return separate fields in application logic. Sequential updates and common-ancestor values are not parallel conflicts. The same conflict rule applies to workflow state.

After configured retries are exhausted, an ordinary branch failure is saved locally. Independent branches and their descendants continue. A node requiring the failed branch fails before executing with GRAPH_DEPENDENCY_FAILED. The root ultimately returns a failed outcome, preserving branch failure descriptors. An application may catch an expected failure and return valid fallback data instead. Optional dependencies are not implicit.

A human interrupt pauses only its node. Other branches keep advancing until they finish or pause. All waiting node journals and child snapshots are checkpointed before the root presents one question through its existing handle. Answering that question resumes only its owning node; other questions remain waiting. The resumed node and its independent successors can advance while a shared node still waits. Questions are presented one at a time, not as a batch; a running sibling still delays publishing a durable suspension.

Whole-execution cancellation reaches all running branches. Execution-limit pauses preserve completed and waiting branches for Continue. Use Redis for restart durability and the existing checkpoint APIs for fork/rollback. External actions still require application idempotency; arbitrary in-flight JavaScript is not crash-recovered.

Parallel graphs must be acyclic: they cannot route back to an earlier graph node. Sequential loops inside a child remain supported. Use useWorkflow rather than transitionTo from parallel branches; branch handoffs are rejected instead of abandoning siblings. Existing sequential workflow semantics are unchanged.

Fork, rollback, and persistence

Use the existing session checkpoint APIs. A fork made while a child is waiting contains the nested waiting chain. Snapshot-backed forks get separate run IDs and resume tokens, so source and fork can answer differently. Rollback restores the selected child checkpoint and pending writes. It does not undo changes already made to your application database.

Use Redis or another durable framework adapter for pauses that must survive server restarts. In-memory state lasts only for the current process. Retention/TTL still applies. Built-in memory and Redis stores atomically consume resume tokens; stale, cancelled, or cross-session requests fail. Client cancellation ends the waiting execution, whereas selecting an application-defined "Cancel" option resumes node logic with that value.

The saved runtime context is retained during resume. New client history or picker context does not silently change the enclosing node's routing. Authorize every request on the server before invoking the agent, including resumed and forked requests.

Limits and migration

Within one node, await calls sequentially or join independent calls with parallel([...]). Native Promise.all child calls are rejected. Acyclic calling/called graphs may have parallel edges.
Use a unique, stable call id within each node activation. Re-entering a node through a graph loop starts a fresh child invocation. The limits are 64 calls per node activation and 16 nested child levels.
Child input and output must be JSON values. Avoid undefined, functions, dates, maps, cycles, and non-finite numbers.
Each child has isolated input, accumulated data, node state, and workflow state. Server runtime context, provider access, and tracing are inherited.
Child streams join the root stream with scoped IDs. Completing a child does not end the parent stream.
transitionTo inside a child is rejected. Convert nested handoffs to awaited child calls and return the fields the caller needs.
Completed child results and failures are cached through parent replay/retry until the parent node commits. Retrying a failed child intentionally requires a distinct stable call ID and a bounded application retry policy.
Change workflow versions when changing suspended execution contracts. A changed typed version or changed active child version fails replay; the runtime does not load historic code automatically.
Durability is at checkpoints. This API does not add background scheduling or arbitrary in-flight JavaScript crash recovery. Serialize concurrent rollback/fork/edit operations at the application/session boundary.

The Canvas example uses this API for creation, brief queries, updates, and saves. The update workflow can call save as a nested child. See interrupts and resume and hooks for the surrounding APIs.

Inspect calls in Studio

See the Studio overview for the child execution tree, child rows, and observed call links.