Skip to main content

Workflows

Chain dynamic agents into multi-step automations with the visual workflow builder. Workflows run on demand from the chat UI, on a cron schedule, or when triggered by an approved agent tool call.

Quick links: Helm Chart Docs · Security — Workflow RBAC

Visual Workflow Builder​

Build workflows directly in the CAIPE web UI without writing code.

  • Drag-and-drop step ordering with per-step agent assignment and prompt
  • Each step runs a named dynamic agent with a custom prompt and optional config overrides
  • Error handling per step: abort, retry (with configurable max attempts and delay), or continue
  • Workflow configs stored in MongoDB; optionally bootstrapped via appConfig.workflow_configs in the Helm chart

Run History and Event Timelines​

Every execution creates a persistent run record.

  • Live status updates: pending → running → waiting_for_input → completed / failed / cancelled
  • Per-step event timeline showing agent responses, tool calls, and errors
  • Artifact storage for files produced during a run
  • Shared run visibility: owners can share a run with a team for collaborative debugging

Human-in-the-Loop (HITL)​

Workflows can pause mid-execution and wait for human input or tool approval.

  • Steps can emit an interrupt with a custom prompt and structured form fields
  • The HITL form renders inline in the chat UI — no page navigation required
  • Subagent HITL is supported: a sub-step can interrupt independently of the parent workflow
  • Resume data is validated and forwarded to the waiting agent step

Scheduling​

Workflows can run on a cron schedule without human intervention.

  • Configure cron expressions per workflow in the UI or via appConfig.workflow_configs
  • Scheduled runs are owned by the service account and visible to team admins
  • Schedule history and next-run time shown in the workflow list

Agent Workflow Tools​

Approved agents can trigger and monitor workflows as built-in tools.

  • Add the workflows built-in tool to any custom agent and grant access to specific workflow configs
  • The agent receives the run ID and can poll for status or surface results inline in chat
  • Webhook triggers: POST to /api/workflow-runs with a valid Bearer token to start a run from external systems

Access Control​

Workflow configs and runs follow CAIPE's RBAC model.

  • Config visibility: private (owner only), team (shared with named teams), or global
  • Run visibility: private by default; owners can share a run with a team via the share dialog
  • OpenFGA enforces per-config and per-run access on every API call
  • Org admins can access any run for troubleshooting regardless of share settings