Flow Control

Beyond agents and operators, betool provides control nodes to structure the logic of a pipeline.

Condition

A condition node routes exchanges to multiple outputs based on a predicate.

Configuration:

• Input slot — what is evaluated (exchange.intent, webhook.payload.type…)
• Branches — for each value (or boolean condition), a named output
• Default branch — mandatory (fail-safe)

Typical use: routing based on the intent detected by a classifier agent, or based on an HTTP status returned by an operator.

Loop

A loop node iterates over a collection. Configuration:

• Collection slot — the list to iterate over
• Iteration variable — the local name for the current element
• Limit — maximum number of iterations (safety guard)

On each iteration, the internal subgraph executes against the current element. Outputs are aggregated into a list.

Filter

A filter node removes exchanges that do not satisfy a predicate. Useful for eliminating noise before incurring downstream LLM costs.

Example: filtering emails whose from field is on a blocklist, or webhooks where type != "order.created".

Human approval (confirmation)

A confirmation node pauses the pipeline and waits for a human to approve or reject from the admin interface.

Configuration:

• Question — what is asked of the reviewer (with context placeholders)
• Branches — approved / rejected (and optionally custom options)
• Timeout — if no response is received, route to the default branch
• Notification — who to alert (email, Mattermost, Slack)

Typical use cases:

• Confirming a high-value outbound call.
• Validating a cancellation email before sending.
• Approving a sensitive response before publication.

The history preserves the reviewer's name and timestamp — fully traceable for audit purposes.

Best practices

• Always wire the default output. An exchange that falls into a void stays in processing and blocks the queue.
• Keep conditions simple. A condition with 12 branches is a signal that you need two cascaded agents rather than condensed logic.
• Limit nested loops. A loop inside a loop inside a loop means costs that explode. Prefer transforming the collection upstream instead.