Orchestrating workflows and agents with Playbooks

A playbook in CloudRadial AutomationAI pursues a goal by running an ordered sequence of steps — each step is a workflow or an agent you already have. A playbook can run once, or keep running on its own and use what it learned last time to do better on the next run. You don't wire loops or draw a flowchart; you answer four plain questions and AutomationAI takes care of the rest. This article is for workspace users who build and operate automations (Admin or Owner Role).

• How a playbook works
• The building blocks
• Building your first playbook
• The five kinds of step
• Running once or repeating
• How a playbook learns
• Watching your playbooks
• Faster starts
• A worked example
• Glossary

How a Playbook Works

A playbook is built from four answers:

• What's the goal? A sentence or two describing what you're trying to achieve.
• What are the steps? The ordered things it should do — run a workflow, run an agent, wait, or check with you.
• When does it run again? Just once, on a schedule, when something arrives, or as soon as it finishes.
• When is it finished? The rules that stop it — goal reached, a run limit, a date, or a spend cap.

A Workflow is one precise automation you build and test on a canvas. A playbook sits a level above: it strings several workflows and agents together toward an outcome and can keep at it over days or weeks. Workflows are the parts; a playbook is the plan.

The Building Blocks

Everything a playbook uses already lives in the left sidebar, in two groups. Actions are the things that do work. Foundations are what they draw on.

• Workflow — a precise, testable automation built in the Designer and run on your runners. The granular piece you set up and prove out once.
• Agent — an AI worker you hand a goal. It reasons, calls tools, and returns a structured result you can use in later steps.
• Knowledge — a folder tree of documents your automations can read and write. Agents can be grounded on it, and it's where a playbook keeps what it learns. Top-level folders double as workspaces.
• Extensions — connections to outside systems and vendor APIs. They give an agent its tools: install the extensions an agent needs and it can act in those systems.

You rarely build these from scratch for a playbook. You point at workflows you've already published and agents you've already installed. When you're ready, open Playbooks in the sidebar.

Building Your First Playbook

Select New playbook on the Playbooks page. The builder is the four questions, top to bottom.

• Goal. Give it a name and write the goal in plain language. Optionally add a headline metric (the number you care about, such as ROAS) that shows on the card, and pick a home workspace — a Knowledge folder where this playbook keeps its notes.
• Steps. Add steps in the order they should happen. To run two steps at the same time, turn on Run at the same time as the step above. They indent together into one group, and the playbook waits for the whole group before moving on. That is the only parallel concept you need; there is no canvas and no arrows to draw.
• Repeat. Choose when it runs again, or leave it on Just once.
• Finish. Select the rules that should end it.

Save it as a Draft while you work. When it's ready, select Start and it becomes Active. You can Pause, Resume, or Stop it at any time. Editing an Active playbook is safe: a run that's already going finishes on the version it started with, and your edits take effect from the next run.

The Five Kinds of Step

Every step in a playbook is one of these. Mix and match them in any order.

Passing Results Between Steps

A step can use what an earlier step produced, the goal, the auto-assembled briefing, or an incoming trigger by referencing it in curly braces:

# the "summary" field from a step named "research"
{{ plays.research.output.summary }}

# the playbook goal, and what it has learned so far
{{ playbook.goal }}   {{ playbook.briefing }}

# a field from the event that started this run
{{ trigger.ticketId }}

A step can only reference steps above it — results that already exist when it runs.

Running Once or Repeating

For When does it run again?, pick one:

• Just once — run the steps a single time (the default)
• On a schedule — hourly, daily, weekly, monthly, or a custom time. If a run is still going when the next one is due, it's skipped and noted.
• When something arrives — each qualifying message to the playbook's trigger link starts a run, queued if one is already going
• Again as soon as it finishes — start the next run when the last ends, optionally after a wait. This is the tight keep-optimizing loop.

Two optional add-ons: Only run if a condition holds, and Check with me before each run to gate every run behind your approval.

For When is it finished?, select any combination — the playbook stops when the first one is met: Goal reached (the Reflect step judges the goal met a few runs in a row), After N runs, By a date, Spend limit (in Usage Credits), or Too many failures in a row.

Any repeating playbook comes with two safety limits switched on: it stops after 25 runs and after 3 failed runs in a row. Raise, lower, or turn them off deliberately, but they're on by default so nothing loops forever unnoticed. You can also Pause or Stop from the card at any time.

How a Playbook Learns

A repeating playbook isn't just a timer. If you gave it a home workspace, it builds up memory on its own with no wiring required:

• Before each run, it assembles a briefing: your goal, what the last run's Reflect step decided to change, and the most relevant notes recalled from the workspace.
• After each run, it writes a journal entry into the workspace — what ran, how it went, and the Reflect verdict.
• The next run recalls that journal automatically, so Run #2 knows what Run #1 learned, Run #3 builds on both, and so on.

That's the whole loop: add a Reflect step, give the playbook a home workspace, and let it repeat. Each pass is shown to you simply as Run #1, Run #2, and so on.

Watching Your Playbooks

The Playbooks home is a live board. Each column is a workspace (one of your top-level Knowledge folders), plus a Company-wide column on the left for playbooks that aren't tied to one workspace. Each card is a playbook, showing its status, its current Run #N and step, its headline metric, and anything waiting on you.

A card's status is shown with both a label and a shape, so it reads clearly at a glance and never relies on color alone:

When a playbook needs a decision, its card reads Needs approval and links to your Inbox, where you approve or deny. Open any playbook to see its run history: every Run #N with a timeline of its steps, links into the underlying workflow or agent run for the detail, the Reflect verdict, and a link to that run's journal entry.

Faster Starts

You don't have to build from a blank page:

• Generate with AI — describe the goal in a sentence and AutomationAI drafts a playbook wired to your real workflows, agents, and workspaces. It arrives as a Draft for you to review and adjust before starting.
• Import a playbook someone shared as a file, and Export yours to share or version. If an imported step points at a workflow you don't have, it still imports as a Draft and flags that step so you can fix it.

A Worked Example

Goal: improve return on ad spend across Google and Meta campaigns; aim for 3.5 and hold it. Home workspace: an Advertising Knowledge folder. Repeat: again as soon as it finishes, after a one-day wait. Finish: goal reached (twice running), after 25 runs, or a 500-credit spend cap.

1. Research the market (agent), running at the same time as Pull competitor ads (workflow)
2. Generate ad variants (agent), automatically briefed on what the last run learned
3. Check with me (approval) — approve the spend before anything goes live
4. Launch ads (workflow)
5. Wait three days (wait) — let the numbers mature
6. Track results (workflow) — save the metrics to the workspace
7. Reflect (reflect) — what worked, what to kill or scale, what to change next run

Run #1 explores. Run #2 opens by reading Run #1's journal and its Reflect notes, and adjusts. Nobody drew a loop — you described the work and let it repeat.

Glossary