Steps

A workflow consists of a trigger and steps. There are three kinds of steps:

• Agent: a step with AI. It reads fields from the context and calls the tools you give it.
• Action: a step without AI. It runs exactly one operation from the catalog, for example Send email.
• Condition: a branch without AI. It checks the value of a field and chooses the path accordingly.

Add a step

1. Click a + on the canvas.
2. In the menu New step, choose the kind: Agent, Action or Condition.
3. Click the new card to configure the step.

How you connect and test cards is described under The builder.

Trigger

Every workflow has exactly one trigger. It is always at the start, has no incoming connection and exactly one outgoing one. It fills the first fields of the context, for an email for example sender, subject and content.

The field Responsibility on the trigger filters out unsuitable events. There is no separate card for this. The kinds of trigger and the responsibility are described under Triggers.

Context

The context is the list of fields of a workflow. The panel Context on the canvas shows all of them. The fields belong to the workflow, not to a step. Data does not flow along the arrows. Each step only sets which fields it reads and which it fills:

• Input: the fields the step reads. The trigger or an earlier step must have filled them before.
• Output: the fields the step fills itself. It only moves on once each of them is filled or marked as No value.

In the run, an agent sees the values of its input and output fields. Of all other fields it sees only the name and the state: filled or still empty. Anything that is not a field in the context is unknown to it.

If an agent cannot determine a promised value, it enters No value with a reason instead of making something up.

On the field you set Multiple entries:

• Replace (default): the newest value applies.
• Collect: every entry is added.

Content from outside

Fields with content from outside carry the marking External. These are email contents, submissions from a public form and data from an integration. The agent receives the marking in the run, together with the rule to treat such content only as data and never to follow instructions in it.

For a trigger that delivers content from outside, lowcloud always sets the marking. This also applies to the fields of a start form on that trigger.

Agent

A click on an agent card opens three tiles: Input, Work and Output. You change the name of the step in the header of the dialog. Below it is the row Note, a remark for you and for the chat in the column Build. The agent never reads the note in the run.

Input

• Context: the fields the step reads.
• Knowledge: documents the agent searches in every run.
• Skills: instructions from the skill library. The step looks one up when it needs it.
• Buckets: shared collections that grow with every run. Each bucket has the switches Read and Write. With Write, the step stores things in it without asking.

Knowledge, Skills and Buckets exist only with Own agent.

Work

• Instruction: the text the agent follows in the run. It says what the agent checks and what it is responsible for. If the instruction is missing, the dialog and the card point this out.
• Optimise text with AI: rewrites your draft. Undo brings it back.
• Model: needed only if this step requires a particular model. Otherwise it stays on Automatic.
• Tools: what the step may do itself, from the catalog. Add one with + Tool. The tools are grouped by service. You choose the account a service runs through in the header of its group.
• Limits > Work steps: how many tool calls and messages to you the step may make per pass. The default is 6. Context maintenance, knowledge search and reading skills and files do not count.
• Questions: if this switch is on, the step may ask you a question during the run and waits for the answer.

The switch Questions exists only on a manual start, because then someone is watching. With email, schedule or a public form, the builder does not offer it. In a run that someone is watching, every step may also send you interim messages and keeps running. This does not depend on the switch.

Next to Instruction you choose Own agent or an agent from the library. An agent from the library brings its own model, tools and knowledge. Model, Tools and Work steps then do not apply. Instead you choose under Accounts which account a service runs through. Per service, one account applies to all steps of the process.

Set fixed values

For each parameter of a tool you choose Agent decides or Fixed value. The agent cannot change a fixed value, such as a recipient, a folder or a channel. If a fixed value stays empty, the card reports it, and the workflow cannot be published.

Output

The tile Output shows under Context the fields the step fills, and under Next step the path after this step.

Action

An action decides nothing and words nothing. It calls the tool with the inputs you define. Use an action when it is fixed which operation should run with which values. If the step has to weigh something up or word something, use an agent.

A click on the card opens the tiles Work and Output:

• Tool: the one operation of the action, from the built-in operations of the catalog. For a different operation you create a new action.
• Accounts: which account the service of the operation runs through.
• Inputs: one value per parameter. This is a fixed value, a field from the context or text with fields. A field on its own is passed on unchanged.
• Result: which value from the response lands in which field of the context. Later steps can then read it.
• Next step: another step or End.

Condition

For a decision, an agent enters a value into a field. The condition after it checks the field and chooses the path. A click on the card opens the tiles Input and Rules:

• Field: the one field the condition checks. You can choose fields of the type Selection, Yes/no, Number or Date. A field that collects entries cannot be checked.
• Rules: each rule consists of a comparison, a value and a target. The comparison depends on the type of the field, for example Is for a selection, Below or At least for a number, Before or After for a date. Is empty exists for every type.
• All others: the last path. It applies when no rule matches.

The rules are read from top to bottom, and the first one that matches applies. Use the arrows on a rule to change the order. End is also possible as the target of every path.

A field that was never filled or is marked as No value counts as empty. Without a rule with Is empty, the run then takes the path All others.

Turn approval on or off

Every tool in an agent and in an action has the switch Approval. If it is on, the run stops before each call and shows you the arguments of the call for confirmation. If it is off, the step runs the tool immediately.

For tools that write or send something, approval is on by default. Reading tools run without asking. If you switch off approval on a writing tool, lowcloud asks you to confirm. After that the tool carries the marking Without approval.

There is no separate card for approvals on the canvas. What an approval shows in the run and how you answer it is described under Approvals and questions.

Paths

The arrows between the cards are the paths. An agent and an action have at most one path:

• Without a path, the process ends after the step. The answer of an agent is then the answer of the run.
• With one path, it always continues there.

Only a condition has several paths.

In the dialog of an agent, you add a path with + Path as long as none is set. The path gets a short name and a target. You can also draw a path on the canvas, from the handle of the path row onto the target card. End is also possible as a target. The run then ends successfully at this point, with the agent's short reason instead of a written answer.

Agents with two or more paths only remain in old workflows. There the agent chooses exactly one path at the end. Neither the dialog nor the chat creates such agents any more.

Loops

If a path points back to an earlier step, a loop arises. The step then runs once more, with the latest state. There is no separate card for loops.

On the condition that sends the run back, the path Round limit appears with the field Rounds. Rounds sets how many rounds the loop runs at most. The default is 3. After the last round, the run takes the path Round limit instead of going back into the loop. Its target is a step outside the loop or End. If the target is set to Abort, the run aborts with an error after the last round.

If a loop has no exit, the card reports it.

In old workflows in which an agent with two or more paths decides the loop, Rounds sits on the agent under Limits.