Blog · 2026-11-11
The relationships between agent and workflow
A workspace holds your agents. Each agent does its work through workflows and direct messages, and each workflow can start four ways: Run now, a schedule, an API call or a webhook.
LoopEngine has a few building blocks: workspaces, agents, workflows and messages. Once you see how they fit together, it's easy to decide where a new job should go. Here's the whole picture on one page.
The whole picture
Read it from the outside in:
- A workspace has many agents.
- An agent has many workflows, and also takes direct messages.
- A workflow can start four ways: Run now, a schedule, an API call or a webhook.
A workspace has many agents
A workspace is your team's LoopEngine. It holds the things every agent shares: the apps you've connected (Gmail, Shopify, Stripe, Slack and others), the AI models and their keys, and settings such as the time zone and who can sign in.
Inside it you create agents, one for each line of work. A support agent answers customers. A marketing agent writes and posts content. A finance agent might chase unpaid invoices. Each agent has its own:
- Instructions: who it is, how it talks and what it must never do.
- Apps: which of the workspace's apps it can use, and which of their tools.
- Permissions: what it may do on its own and what it must ask you about first.
- Messaging: where its questions, approvals and results reach you, such as Slack, email or Lark.
Keeping agents separate keeps their access separate. Your marketing agent never needs to see refunds, and your support agent never needs to publish blog posts.
An agent has many workflows, and takes direct messages
An agent gets its work in two ways.
Workflows are jobs it does again and again, the same way each time. "Every 10 minutes, answer new customer emails." "When a Stripe dispute comes in, gather the evidence and ask me before sending it." A workflow saves:
- The steps, in order, with where to stop and ask you.
- What it's given each run, such as an order number or a dispute ID.
- The permissions its tools need, so sending, paying or publishing waits for your approval.
- How it starts, described in the next section.
Direct messages are everything else: one-off requests you send the agent yourself. You can write to it in Chat in LoopEngine, message it in Slack, Telegram or email, or send it a message from your own app. "Where is order #1001?" doesn't need a workflow. The agent answers with the same instructions, apps and permissions it uses for everything else.
Every workflow belongs to one agent. When you build a workflow, you pick the agent first. LoopEngine plans it with what that agent already has, and tells you which apps or messaging the agent still needs before you create it. If the job doesn't look like that agent's kind of work, it says so and suggests creating a new agent instead.
A workflow starts four ways
Run now. Press Run now on the agent's Workflows tab to try a workflow once, or ask the agent in Chat to do it. This is the best way to check a new workflow before it runs on its own.
Schedule. The workflow runs at a time or on an interval: every weekday at 8:00, every hour, the first of the month. A new schedule starts switched off, so nothing runs until you turn it on.
API. Your own software starts a run with one request, and passes the details the workflow needs:
POST /agents/cs-agent/workflows/customer-reply-draft/run
{ "params": { "order_number": "EM-1001" }, "wait": true }
With "wait": true the answer comes back in the same request. Without it, you get a run ID right away and can check the result later, or have it sent to a callback address.
Webhook. Many services can notify a web address the moment something happens: Stripe when a dispute is opened, GitHub when an issue is filed, Shopify when an order is paid. Each workflow can have its own webhook address. LoopEngine checks that the request really came from that service, picks the details out of it (the dispute ID, the order number), and starts the run.
However a run starts, it follows the same steps and the same rules. If a step needs your approval, the run waits and asks you, in Chat if you're there, otherwise through the agent's messaging. It waits for minutes or days, and nothing is lost in between.
Which one do you need?
A quick guide for your next job:
- A one-off request? Send the agent a direct message.
- The same job again and again? Build a workflow on the agent that does that kind of work.
- A different line of work, or different access? Create a new agent, then give it workflows.
Most teams end up with a handful of agents, each with a few workflows, and send them direct messages in between.