Port a workflow to cloud agents

A workflow is a set of configured steps Harriet runs for you. A cloud agent task is a set of instructions an agent follows in your cloud workspace, where it has a shell, files and your connectors. Porting hands an existing workflow's configuration to that agent and asks it to rebuild the job as a scheduled task.

Before you start

Starting a port

Open the workflow you want to move, then choose Port to cloud agents from the ⋯ menu beside Run workflow. A new tab opens. If you own more than one cloud workspace, pick which one the work should go to; otherwise Harriet uses the one you have.

Press Port workflow. Harriet copies the workflow's full configuration into a new conversation in that workspace and wakes it if it is asleep. Waking can take a couple of minutes, and you can close the tab: the message is delivered whenever the workspace next starts.

What the agent does

The agent is asked to check whether it actually has the tools to do what the workflow does, then plan the rebuild before touching anything. Read-only investigation it will just do. Anything that would change data, send a message or write to a connected system, it asks you about first.

When you are both happy with the plan, it saves a scheduled task with the same schedule and triggers, and reports back with a link. Talk to it in the conversation exactly as you would about any other piece of work, and it will adjust.

The new task starts paused

Your original workflow keeps running and is not modified. The new cloud task is created paused, so the same ticket or the same Monday morning cannot set both of them off. When you are satisfied the cloud task does the job, turn the original workflow off and then switch the cloud task on.

What does not carry across

The agent never receives the workflow's webhook secret.

Seeing what a cloud task will do

Open a cloud task and choose View flow diagram from its ⋯ menu, or press Flow diagram on the task in Scheduled tasks. Harriet reads the task's instructions and draws the flow it expects: how the task starts, the work it implies, where it would stop to ask you something, and what it produces. It is a picture of the plan, not a record of a particular run, so use the run history for what actually happened.

The diagram is redrawn whenever the task is saved. Workspace viewers can see it too.

Your organization's skills that the instructions name, such as Load the weekly-report skill, are listed under the instructions, with a link to each skill you can open. Skills that exist only as files in the workspace are not listed. A skill switched off for the workspace is flagged there, because the task cannot use it until you turn it back on under the workspace's Skills settings.

Going back to earlier instructions

Choose Instruction history from the task's ⋯ menu to see every saved version of its instructions, newest first, including changes the agent made from its conversation. Select a version to see what restoring it would add and remove, then choose Restore this version. The restored instructions are saved as a new version, so you can undo a restore the same way. The schedule, triggers and model are left as they are.

Only the workspace owner can restore a version. Viewers can read the history.

Guardrails