Agent DocsDeveloper documentation
Builder Guide

Overview

What a workflow is, the SDK as a thin wrapper over Temporal, workflow vs agent

Workflow Overview

Audience: Customer developers building automation on the Agentic Hub, and admins who govern it. You write your workflows in your own repository; the platform runs and governs them.

A workflow is durable, long-running automation: a sequence of steps that can call agents and tools, pause for a human decision, wait days for an email reply, retry on failure, and survive a process crash without losing its place. You write workflows in Python with the Workflow SDK (workflow_sdk) in your customer repository, and the platform runs each one on an isolated, governed worker.

Workflow vs. agent

Use the right tool for the job:

AgentWorkflow
ShapeConversational / single-turn reasoningMulti-step, long-running process
DurationSecondsSeconds to weeks (waits for humans, emails, timers)
StateMostly stateless per turnDurable, survives crashes and restarts
Best for"Answer this", "summarise that""Onboard this vendor", "review and approve this case"

Workflows and agents compose: a workflow orchestrates the process and calls agents for the reasoning-heavy steps (see Calling agents & tools).

A thin wrapper over Temporal

The Workflow SDK is a thin abstraction over Temporal. Temporal is the durable-execution engine that gives your workflow its superpowers — durable state, automatic retries, timers, and crash recovery. The SDK adds the agentic-hub patterns on top:

  • Correction loops — repeat a step with human feedback until it's right.
  • Human-in-the-loop (HITL) — pause for a governed human review.
  • Email watcher — park the workflow until a matching reply arrives.
  • Managed state — typed, immutable, auditable Pydantic entities.
Scroll to zoom · Drag to pan · 100%

You can drop down to raw Temporal. The SDK re-exports Temporal's own workflow and activity, so you can call workflow.execute_activity, start child workflows, set timers, and define your own signals directly. But then the agentic-hub patterns above do not come for free — HITL, correction loops, and the email watcher follow agentic-hub implementation details you'd have to reproduce yourself. Start with the SDK patterns; reach for raw Temporal only when you need something they don't cover.

What the platform runs it on

You never operate a runtime. You edit code, commit, and the platform builds your code and rolls it onto a personal, isolated remote worker — one per developer. That inner loop is covered in The dev loop & remote workers.

Governance, from the start

Everything a workflow does that touches the outside world is governed:

  • Calling an agent requires an explicit grant (you declare it; an admin approves the policy).
  • Human review is a first-class, auditable step — you declare exactly what a reviewer sees.
  • State is immutable and append-only, so every decision can be reasoned about and audited.

This section

  1. Overview — you are here.
  2. The dev loop & remote workers — how your code becomes a running worker.
  3. Building a workflow — the registration contract, steps, and correction loops.
  4. Human-in-the-loop — governed human review, and when to use it.
  5. State & the entity model — typed, immutable, auditable state.
  6. Calling agents & tools — invoking agents under policy.