Darwin Academy · 06

Processes — I work while you don't

A client's files land in a folder, and tasks appear on your list. A mail arrives, and it's already handled when you look. That's a process: a named sequence of steps that one of my agents runs — on a schedule, on an event, or on your word. This is the complete guide to setting them up.

By Darwin · the voice AI assistant that runs on your own PC

What a process actually is

A process is a short brief, written the way you'd brief a person: a name, an agent who owns it, and steps in plain language — check this, read that, do this with it, tell me the result. No flowcharts, no programming. The steps are instructions for the agent's brain, so “read the attached offer and create tasks from what it actually asks for” is a perfectly good step.

One rule keeps it all manageable: one process = one thing. Many small pieces beat one giant macro — small processes are easy to reason about, easy to chain, and easy to switch off.

Every process has an owner agent. The owner runs the steps and — this matters — writes what it learned into its own memory, so the process gets better with every run. Your invoicing process remembers the quirks of your invoices.

Creating one — the window, or just say it

Open Processes on the dashboard: a form with the name, the trigger, the steps, and how you want to be told about the result. Or skip the form entirely and tell me what you want:

“Every workday at 7:30, check the mail and read me what matters.” · “When files land in the Novák folder, go through them and propose tasks.” · “Remind me in half an hour to call the bank.”

That last one shows a nice detail: a one-off reminder is just a process with a run-once trigger — and after it fires, it deletes itself. No graveyard of dead reminders.

Triggers — when a process runs

The scheduler checks triggers about once a minute — so “instant” means within a minute or two, which is exactly what you want from a quiet background worker.

The watched folder — my favourite

The scenario this was built for: a client's documents arrive, get saved into the order's folder, and the process takes it from there — reads them, proposes what to do, creates the tasks. The mail is just the doorbell; the work comes from what's actually in the files.

So the process doesn't get a vague “something changed”. It gets the list: new folder “Novák”, new file ponuka.pdf (1.2 MB) — and instructions to work with those exact files. A few thoughtful details you don't have to configure but will appreciate:

• Half-copied files don't count. A big file still being copied is left alone until it's been quiet for a while — the process never runs on half a video.
• Batching. Tell it to wait for at least N new items, and it gets them all at once instead of one by one.
• Folders come and go. Orders start and finish — you add and remove watched folders on the fly (“watch the Atreon folder too”), without touching the process itself.
• An unplugged disk isn't an event. If a path is briefly unreachable, nothing fires on the false “everything disappeared”.

Results — how you hear about it

Per process, pick any mix of: voice (I say it), desktop notification, phone, e-mail, or just a quiet log. If you have Telegram or PrivateChat connected, those work as channels too. And any process can report into the morning brief, so the night's work greets you with your coffee.

A process can also carry an approval step — do the work, but ask before the part that matters. The question reaches you on the dashboard and your phone; yes proceeds, no stops, and a written note sends the agent back to revise.

Chaining — small pieces, bigger machines

Because one process is one thing, real workflows are chains: a process can emit a named signal when it finishes, and other processes — yours, or a teammate's Darwin in Team mode — can be triggered by it. Documents-in fires reading; reading done fires task creation; tasks created pings your phone. Each link stays small enough to understand.

One built-in worth knowing: commands by e-mail

Darwin ships with one prepared process, switched off by default: commands by e-mail. Turn it on, and a mail from you with a subject starting Darwin: becomes a direct instruction — I do the work and reply to your address. Sensitive actions still demand extra verification, so a spoofed mail can't puppet me.

Who can change your automation — the safety model

• Creating, editing or deleting a process sits behind your access code. Running an existing one is open; changing what runs unattended is not.
• A running process cannot rewrite the automation. Unattended runs are locked out of creating or modifying processes entirely — an instruction smuggled into a mail or a document can't plant a permanent worker on your machine.
• Pause beats delete. A paused process keeps its memory of the watched folders, so when you re-enable it, it knows what arrived in the meantime.

Set it up once, in plain words — and a piece of your work simply stops being yours to remember.

What's next

• Get the results on your phone — connect Telegram › or the encrypted PrivateChat line ›
• Let processes watch your company mail — connect Outlook / Microsoft 365 ›
• Meet the owners — your team of agents ›