Routines

A routine is work an agent does on its own, on a schedule: a stock check every weekday morning, a look at open pull requests every hour, a weekly summary for a client. You hear back only when there is something to say.

Routines live in each project’s Routines tab (Home’s are in the base project). New routine opens the editor.

APX

Routines, with their last activations

Actions

Each routine does one of four things:

Action	What happens
Agent	An agent (Home or one of the project’s) answers your prompt. It tells you only when it has something to say.
Message	Sends you a fixed message on your channel (Telegram, WhatsApp…). No model.
Shell	Only runs a command in the project’s folder. No model, no notice.
Watcher	Sweeps your work for what is worth a word (late tasks, commitments coming due, quiet projects) and calls the model only when something is new.
Team news	Tells you in one message what your other agents’ routines and notices had for you; while one is on, they don’t write to you on their own. Calls the model only when there is news.

Schedules

When	Example
On a schedule	Pick days and time (“Weekdays at 08:30”), or Write a cron: five fields, in your local time. 30 8 * * 1-5 is weekdays at 08:30.
Every	Every N minutes, hours, days or weeks. At least 5 minutes.
Once	A single date and time.
By hand	Only when you press Run now.

A run missed while the core was off runs once when it comes back, never twice.

Silence is an answer: SKIP

An agent routine’s prompt is your own words. When there is nothing worth telling you this time, the agent answers only SKIP and nothing reaches you. Write prompts that leave room for that:

Check yesterday’s orders for anything unusual: refunds over 200, a product selling 3× its average, a failed payment. If nothing stands out, answer SKIP.

Commands before and after

Add commands before and after gives a routine a pre-command and a post-command. Both are shell commands you write (agents cannot), run in the project’s folder, two minutes at most and without credentials in their environment.

1. Pre-command runs first. Its output reaches the agent where you write {{pre_output}} in the prompt (or after the prompt if you don’t), marked as outside data.

2. Call the model decides whether the agent runs, from the pre-command’s result:

   Option	The model is called
   Always	every time
   If it printed something	only when the pre-command printed output
   If it failed	only when the pre-command exited with an error
   Never	never (the commands run alone)

3. Post-command runs at the end with what came out, in these environment variables:

   Variable	Value
   APX_LLM_OUTPUT	What the agent answered (or the message)
   APX_STATUS	completed, failed or skipped
   APX_SKIPPED	1 when the model was not called, else 0
   APX_PRE_OUTPUT	The pre-command’s output
   APX_PRE_EXIT	The pre-command’s exit code

That is how a routine spends 0 tokens on quiet days: Tomás’s pre-command lists the low-stock items, with If it printed something, and the model is called only on the days there are any.

A Message routine uses {{pre_output}} too: “Today’s weather: {{pre_output}}” sends the command’s output as it is, with no model at all.

Anchors: always arrives

Routine news normally goes through your interruption budget: a daily cap, gaps, quiet hours. Turn on Always arrives for a routine you never want held back, like a morning brief: its news arrives even at night or past the cap, and does not count against it.

Run now and the activation bars

Run now starts a routine immediately, whatever its schedule. Each routine shows its last activations as 14 bars, oldest to newest:

Bar	Meaning
The action ran	It ran and had nothing to tell you.
Told you	It ran and sent you something.
Skipped, 0 tokens	The model was not called (the pre-command did not ask for it, or nothing new to watch).
Failed	The run failed.
Did not run	A time it should have run and did not.

Each run is an ordinary chat of its own, so you can open it and see exactly what the agent did.

From the terminal

apx

❯ apx routines help
apx routines                                   Routines and when they run next
apx routines add <name> <prompt…> --agent a [--project p] (--cron "30 8 * * 1-5" | --every 60 | --at <ISO date>)
               [--pre "<command>"] [--call always|output|fails|never] [--post "<command>"]
                                              The pre-command's output reaches the prompt at {{pre_output}}
apx routines run <id>                          Run now
apx routines on|off <id>
apx routines rm <id>
apx routines log <id>                          Recent runs

With no --cron, --every or --at, the routine runs by hand. --every is in minutes. The CLI creates agent routines; message, shell and watcher routines and anchors are set in the panel.

# Weekdays at 9: Tomás looks at low stock, only when there is any
apx routines add low-stock "Here is today's low-stock list: {{pre_output}} Draft the reorder for the supplier." \
  --agent <tomas-id> --project <acme-id> --cron "0 9 * * 1-5" \
  --pre "./scripts/low-stock.sh" --call output

apx routines run <routine-id>     # try it now
apx routines log <routine-id>     # recent runs; "(nothing to say)" marks a SKIP

Note

Routines are yours: only you create or change them. An agent cannot schedule standing work for itself.

Next

SecretaryThe day open, day close and watch routines, and your interruption budget.

ExtensionsExtensions can suggest routine templates.