Programs

A Program is a scheduled check-in the agent starts. On a cadence you set, it sends each enrolled person an opener, a Meta-approved template on WhatsApp or a message you write on Slack, waits for their reply, runs a skill with them when they answer, and records how it went. Checking in with a group of people on a schedule, morning and evening, is the case it was built for.

A Program is five things: a skill (the questions), a cadence (which weekdays, at which local times, in which timezone), a channel and sender, an opener (the message that starts the conversation), and a roster of users. Ops holds that configuration; the agent runtime does the sending and keeps the per-user record.

Open it from an agent’s left navigation (Programs, under the agent) or go to /work/agents/<agentId>/programs. The entry appears only when the Programs capability is on for that agent.

Programs run on WhatsApp and Slack, the two channels where the agent can start a conversation. Telegram cannot be used because a bot may not write first there, and the website widget has no address to write to.

Before you start

Five things have to be in place, and the activation gate checks every one of them:

A connected channel — WhatsApp with its Business Account ID filled in, or Slack in Personal assistant mode; see WhatsApp and Programs on Slack.
An opener. On WhatsApp, an approved template: register it under Publish → Channels → WhatsApp → Manage → Templates and wait for Meta to approve it; see Templates. On Slack, a sentence you write on the Program; nothing to approve. Either way keep it short: Time for your check-in. Reply to begin. The questions do not go in it.
Users bound and permitted — a phone number for WhatsApp, a linked Slack member for Slack, and Permission to message them first; see Users.
A skill with the questions — an ordinary skill available to the agent. The agent runs it when the user replies, so write it as the conversation you want held: what to ask, in what order, what counts as done, and when to escalate.
The capability on — Configure → Capabilities → Programs. It is off by default because a Program messages real people on a schedule.

A skill edit reaches the agent when the agent’s bundle is republished, not the moment it is saved. Check the bundle before assuming a corrected question is live.

Creating a Program

Click New Program and fill in the form. Saving creates a draft; nothing is sent until you activate.

Field	Notes
Name	For your own list, e.g. Post-op daily check-in.
Skill	The skill reference the agent runs on reply, e.g. `post-op-checkin`.
Channel	WhatsApp or Slack, among the channels this agent has connected.
Sender	Filled in from the connection: the WhatsApp number’s id, or the Slack app.
Template (the opener)	WhatsApp only. The approved template’s name, e.g. `daily_checkin`.
Template language	WhatsApp only. The template’s language code, e.g. `en_US`. A template is approved per language; name and language have to match what Meta shows.
Opener message	Slack only. The text the bot posts to start the check-in.
Days	The weekdays it fires.
Times of day	One or more local times, `HH:MM` on a 24-hour clock, e.g. `09:00, 21:00`.
Timezone	A named zone such as `Europe/Bucharest`. An offset like `+03:00` is refused: it cannot follow daylight-saving changes.
Reply deadline (hours)	How long after the opener a user has to reply before the check-in is recorded as unanswered. Default 24.

The cadence is a set of weekdays and local times, never an interval. Two times on the same day are two separate check-ins; the row shows the next one, in the Program’s own zone, so twice-daily Programs show the next time, not the next day.

Users

Open a Program to see its Users section: every user on the agent, with a checkbox for membership and, beside each name, whether they can be reached on this Program’s channel — a phone number and permission for WhatsApp, a linked Slack member and permission for Slack. Tick to add, untick to remove. Someone who cannot be reached can still be ticked; the activation gate names them.

The roster is the denominator every run counts against: a user who cannot be reached on a given run stays on it, marked skipped, rather than disappearing from the count. An active Program keeps at least one user; pause it to remove the last.

Activating

Activate either turns the Program on or refuses and tells you everything that is wrong at once, so you can fix it in one pass rather than one save at a time. The blockers it can name:

the Programs capability is off for this agent;
no weekday, no time, an unparseable time, or an offset instead of a named timezone;
no skill, or no sender number;
on WhatsApp, no template, or a template name without its language;
the template is not on this WhatsApp account, or Meta reports it as anything other than approved;
on Slack, no opener message, Slack not connected for that sender, or its channel in Team member mode;
the roster is empty;
a user on the roster has no phone number (WhatsApp) or no linked Slack member (Slack), has not given permission, or has since left the Slack workspace.

Whoever activates a Program becomes its responsible operator: escalations from its check-ins land in that person’s Inbox.

Once active, the row shows the next run in the Program’s timezone. Edits that would break the gate on an active Program are refused; pause it first. Removing the last user from an active Program is refused for the same reason.

What happens on a run

At each slot the runtime opens an occurrence and writes one row per user on the roster.

Skipped users are recorded first. A user whose permission has been withdrawn, whose number is gone, or whose previous check-in is still open (Previous check-in still open) gets a row with that reason and is not messaged.
The opener goes out to everyone else — the approved template on WhatsApp, or the opener message as a direct message from the bot on Slack — through the same delivery path as the agent’s replies. Permission is re-checked at the moment of sending, so a withdrawal between the slot and the send still holds.
The reply deadline starts when the opener is handed over. Meta’s delivery reports move the row from Queued to Sent and Delivered, or to Failed with Meta’s reason. Slack confirms the post and sends no delivery receipts, so a Slack row stops at Sent.
When the user replies, the message lands in their own conversation with the agent — the WhatsApp chat, or the Slack direct message — and the agent is told to run the Program’s skill with them now. Only a first reply within the deadline counts; the conversation then continues as an ordinary chat, on WhatsApp inside Meta’s 24-hour window.
The agent closes the check-in with checkin_outcome — Completed when it got the answers, Declined when the person refused or asked to be left alone. If something in the answers needs a human, it calls checkin_escalate, which raises Check-in needs your attention in the responsible operator’s Inbox without ending the check-in.
The deadline sweep records what never finished. A delivered opener nobody answered becomes No reply by deadline. A reply the agent never closed becomes Answered, not finished. An opener that never reached the user goes back to Not started, so one failed send never blocks that user’s next check-in.

A user has at most one open check-in per agent. If a Program is paused, its unsent openers are withdrawn; openers already handed to Meta cannot be recalled. After an outage, a slot whose reply window has already passed is skipped rather than sent late: nobody receives yesterday’s “good morning” this afternoon.

Reading the results

Every row on the Programs page carries the outcome of its last run in one line, so a broken Program is visible without opening anything:

Sent to 12 of 14. 2 skipped: no permission to message. 7 replied; 5 no reply by the deadline. 1 escalated.

A line that reads Sent to 0 of 14. 14 not yet sent. an hour after the slot means the openers are stuck, not that nobody answered.

Click a Program’s name to open its run history: a chip per occurrence, newest first, and for the selected one a table with a row per user.

Delivery, response and escalation are three separate columns because they are three separate facts. A failed send must never read as a user who stayed silent.

Column	Values
Delivery	Queued, Sent, Delivered, Failed, Undelivered, Skipped, Cancelled
Response	Not started, Awaiting reply, Replied, Completed, Declined, Answered, not finished, No reply by deadline
Escalation	— , Escalated, Escalation failed
Reason	Why a user was skipped or cancelled, or Meta’s error text for a failed send

The filters above the table (Everyone, Failures, Skipped users, No reply, Escalated) narrow the rows; the counts in their labels are over the whole roster.

Pausing, resuming and deleting

Pause stops new openers on the runtime’s next scheduling pass and withdraws any not yet sent. History stays.
Resume (Activate on a paused Program) re-runs the gate and starts at the next future slot. Missed slots are not sent.
Delete asks for confirmation, then removes the Program and its roster. Past runs stay in the runtime’s records but are no longer reachable from the page.
Turning the Programs capability off is refused while any Program on the agent is active. Pause them first; otherwise the runtime would stop running them while the page still said Active.

Limits in this release

One opener per Program: a fixed-text template in one language on WhatsApp, one message on Slack.
No reminder after a missed check-in. The next slot simply comes round.
While your Meta app is in development mode, Meta delivers only to the numbers on its recipient list, so a Program appears to work for you and fail for everyone else. That is Meta’s restriction, not the gate’s.

What’s next

WhatsApp for the connection and the Templates screen.
Users for enrolling users by phone or linking their Slack member.
Skills for writing the questions.
Inbox for where escalations arrive.