Orchestrate a workflow from the message box

Most agent work isn't one instruction — it's a small sequence. Clarify what I actually want, then fix it, then have something independent review the fix. Until now that was three messages, sent one at a time, each waiting on the last. As of this release you can write the whole thing in one message: start lines with / or @ and Cockpit reads them as an ordered workflow.

One message, several steps

Every line that starts with a known command becomes a step. The first character picks where it runs:

So this:

Here is the failing test output: payment webhook 500s on retries.
/fx
figure out why the idempotency key isn't being honored
@cr
audit the fix for race conditions and missing rollbacks

…becomes a single numbered plan the AI works through in order. The text before the first command rides along as shared context — paste your log or state the goal once, up top, for the whole run. Everything under a command line, including blank lines and multiple paragraphs, belongs to that step.

You wrote four lines. The agent received a structured plan.

/ keeps it close, @ sends it away

The two markers are about attention, not just routing:

• /verb keeps the work in the current chat, where you can watch it turn by turn and steer. Use it for the steps you care about.
• @verb hands a self-contained chunk — a review, an exploration, a focused investigation — to a sub-agent, which does it and summarizes back without cluttering the main thread.

The everyday shape is "do it here, then send someone to check it":

/go
implement the retry backoff described in the ticket
@cr
review what was just written for correctness and style

Your own skills, in the mix

Steps aren't limited to the built-ins. Any skill you've installed is a verb too, and built-ins and your skills can sit side by side in the same workflow — they all resolve through the same "read this SKILL.md" path. If a skill you wrote shares a name with a built-in, yours wins, so your version is what runs.

And autocomplete now follows your cursor: type / or @ at the start of any line — the second, the third — and the command menu pops for that line. That's what makes stacking steps feel natural instead of fiddly.

Nothing changes for one-offs

If your message is a single /verb with no preamble and no @, it behaves exactly as before — one command, one turn, no ceremony. The numbered plan only appears when there's a real workflow to run: two or more commands, any @ step, or leading context text. The simple case stays simple; the multi-step case finally gets to be one message.

Try it

Update Cockpit, open a chat, and write two commands on two lines — say /qa then @cr. Watch the agent receive them as an ordered plan. Full details in the Workflows docs.

Try it: npm i -g @surething/cockpit · GitHub · Try Online