Core Concepts

Use this page alongside Architecture for the platform-wide mental model—here we focus on definitions vs runs and how work moves.

Cascades models automation as directed workflows: ordered tasks (nodes) connected by edges that express dependencies.

Execution flow (mental model)

This diagram summarizes how authoring becomes auditable outcomes—stop here first when onboarding a new teammate.

Workflow — authored graph with typed nodes (tasks), policies (retries, timeouts), and control flow (workflow builder). Tasks are the units dispatched by workers. Dependencies constrain waves—downstream nodes wait for upstream success unless a conditional route says otherwise. The execution engine applies ordering, backoff, isolation, and artifact capture. Proof generation freezes outcomes into canonical payloads suitable for auditors. Verification compares snapshots vs live definitions, transparency pointers, optional anchors—not just “trust the UI.”

Verification flow (mental model)

The path from frozen definition intent → immutable evidence:

Operational teams live on the upper half (definition → snapshot → execution → proofs). Compliance and platform engineering extend into transparency, anchoring, and future replay for reproducibility—see Replay.

Workflows & DAGs

A workflow definition describes a DAG (directed acyclic graph): nodes hold type, configuration, retries, timeouts, and optional branching. Edges declare which tasks must succeed before downstream work runs—see DAG execution for ordering details.

The workflow catalog lists canonical templates aligned to incident response, billing, proofs, and change control.

Tasks

Each logical unit of work is a task mapped to execution adapters (agents, APIs, integrations, webhook handlers, transforms). Inputs and outputs propagate along edges; failed tasks may retry according to retry policies.

Dependencies

Dependencies constrain execution waves: dependents wait until predecessors complete successfully (or conditional branches evaluate). Cycles are rejected at validation time—see Workflow Lifecycle.

Execution states

Runs and tasks advance through deterministic states (queued, running, completed, failed, and terminal proof states suitable for dashboards and audit tooling). Operational visibility aggregates attempts, backoff behavior, locks, and failure recovery—see Failure recovery.

Integrations touchpoint

Operational systems connect via integration surfaces and webhooks—asynchronous events can enqueue or enrich runs (integration overview).

Related

• Quickstart — run paths end-to-end in a sandbox
• Security overview — execution controls & audit stance