Runs

A run is one execution of a workflow. Every trigger — a click, a schedule, a webhook, a GitHub or Slack event — creates a run, and every run keeps a complete record of what happened.

Lifecycle

1. Start. The trigger's payload and the run's inputs are recorded, and the run is pinned to the current version of the workflow. Later edits don't affect it.
2. Steps execute one at a time in dependency order. For each step Murmurator checks its dependencies and condition, fills in its templates, performs it, and records the result.
3. Finish. The run succeeds when no step failed. If any step failed, the run fails and its summary names the failed steps.

Watching a run

Open a run from the workflow page, the Runs page or the dashboard. Everything updates live, without refreshing:

• the steps diagram colours each step by status as it progresses;
• each step shows its input (after templates were filled in), output, logs and duration;
• llm and agent steps stream the model's answer as it's generated, and list each tool call an agent makes;
• javascript steps show anything written with console.log;
• the trigger payload and the exact version used are shown alongside.

Run lists on the dashboard, the Runs page and each workflow update in real time too, so new runs appear the moment they start.

Failures

A step fails when its tool returns an error, its model call fails, its JavaScript throws or returns the wrong shape, a spend limit is reached, or an agent exceeds max_iterations. The error is shown on the step and summarized on the run. Steps that depend on it are skipped; independent steps still run.

To fix a failing workflow, ask the assistant "why did the last run fail?" — it reads the run's steps, inputs, errors and logs, explains the cause, and can change the workflow for you. Then press Run again.

Runs are resilient to restarts. A run checkpoints after every step, and an agent step also checkpoints after every turn of its tool loop, so when Murmurator restarts the run carries on from the last checkpoint instead of starting over. A finished step is never run twice, and a resumed agent keeps the conversation and the tool results it already had.

If a worker is killed outright, mid-step, the step is run again when it is safe to repeat — JavaScript, a model call, or a tool that only reads. A step that may already have changed something outside Murmurator, such as posting a message or writing to a database, is marked failed with an explanation instead, since repeating it could do the work twice. Press Run again once you have checked.

Cancelling

Press Cancel on a running run. Steps that haven't started are skipped and the run is marked cancelled. A step already executing finishes its current call, but nothing after it runs.

Tokens and costs

Every llm and agent step records its input and output tokens, and the run shows the totals, so you can see what each workflow uses.

• Steps using your own model keys are billed by your provider; Murmurator adds nothing.
• Steps using Murmurator AI models are metered per token. Each model response is recorded on AI models → Usage & limits, linked back to the run. If the account's monthly spend limit or trial credit runs out mid-run, the step in progress fails with a clear message. Steps that depend on it are skipped, and any later Murmurator AI steps fail the same way; steps that don't use a model still run.

See AI models and Billing & usage.

Starting runs

Runs can't start while the account's automation is paused; runs already in progress still finish.