Cortex Workflows

Preview — Cortex Workflows is available on all plans. The API and workflow schema may change during the preview period.

YAML-driven intelligent automation built on the Cortex runtime. Define multi-step workflows that combine LLM reasoning, integration actions, Guard scanning, and Vault audit logging into repeatable pipelines.

Plan requirement: Pro plan and above.

Plan	Active Workflows	Daily Executions
Free	--	--
Pro	10	100
Max	Unlimited	Unlimited
Enterprise	Unlimited	Unlimited

---

Workflow Definition

Workflows are defined in YAML and stored as JSON in the database. Each workflow consists of metadata, an optional trigger, and a list of steps.

name: content-review-pipeline
description: Review and classify incoming content
trigger:
  type: event
  event_type: agent.event.ingested
steps:
  - id: classify
    type: llm
    config:
      prompt: "Classify this content: {{ trigger.payload.text }}"
      model: auto
  - id: check-pii
    type: guard_scan
    config:
      text: "{{ trigger.payload.text }}"
  - id: route
    type: condition
    config:
      expression: "{{ steps.classify.output.category == 'sensitive' }}"
      if_true: escalate
      if_false: approve
  - id: escalate
    type: human_review
    config:
      message: "Content flagged as sensitive"
  - id: approve
    type: vault_log
    config:
      event_type: content.approved

---

Step Types

Type	Description
llm	Send a prompt to an LLM via Cortex routing. Supports model: auto for complexity-based routing.
adapter	Call an external integration (Google Calendar, Slack, HubSpot, etc.) via the integration registry.
condition	Branch execution based on a Jinja2 expression.
transform	Transform data using Jinja2 templates.
wait	Pause execution for a duration or until an external event.
human_review	Pause for human approval. Resumes via the API or dashboard.
webhook_out	Send an HTTP request to an external URL.
guard_scan	Run Guard PII/injection/content scanning on text.
score	Compute or retrieve a trust score for an agent.
knowledge	Query the agent's knowledge base using RAG retrieval.
vault_log	Record a custom event to the Vault ledger.
platform_query	Query VeriSwarm platform data (agents, scores, events).
send_email	Send an email via the platform email service.
web_search	Perform a web search and return results.

---

Triggers

Type	Description
cron	Execute on a schedule (standard cron syntax). Polled every 10 seconds.
webhook	Execute when an HTTP request hits the workflow's webhook URL. HMAC signature verification.
event	Execute when a specific VeriSwarm event type occurs (e.g., agent.scored).
manual	Execute on demand via API call or dashboard button.

---

Variable Interpolation

Steps can reference outputs from previous steps, trigger payloads, and environment variables using {{ }} template syntax.

Variable	Description
{{ trigger.payload }}	The trigger event or webhook body
{{ steps.<step_id>.output }}	Output from a previous step
{{ env.<VAR_NAME> }}	Environment variable

---

Budget Enforcement

Each workflow execution tracks LLM token usage and cost. Per-step budget limits can be set in the workflow definition. If a step exceeds its budget, execution halts and the step is marked as budget_exceeded.

---

Execution Status

Status	Description
pending	Queued for execution
running	Currently executing
waiting	Paused (wait step or human review)
completed	All steps finished
failed	A step encountered an error
budget_exceeded	Token/cost budget exhausted

---

Vault Integration

When Vault is enabled, every step execution is recorded in the ledger with the step type, input/output sizes, duration, and status. This provides a complete audit trail for workflow automation.

---

API Endpoints

Method	Path	Description
POST	/v1/workflows	Create a workflow
GET	/v1/workflows	List workflows
GET	/v1/workflows/{id}	Get workflow details
PUT	/v1/workflows/{id}	Update a workflow
DELETE	/v1/workflows/{id}	Delete a workflow
POST	/v1/workflows/{id}/execute	Trigger manual execution
GET	/v1/workflows/{id}/executions	List executions
POST	/v1/workflows/{id}/executions/{exec_id}/approve	Approve a human review step

All endpoints require authentication via x-account-access-token.

For full endpoint details and rate limits, see the API reference.