⚙️ Processes: AMBA presses the buttons for you, on a schedule

A process is a schedule that presses AMBA's own buttons for you, in order: fetch files from your sources, check whether everything is in, send a reminder, ask an agent to summarize, send the result. Each step is exactly something you could do by hand; a Check step adds "if … then …" branches. You put the steps together yourself (the main way), or describe the process in one sentence and AMBA drafts the steps for you to review. The AI model only drafts; it never turns a process on.

Processes belong to the project; agents do the Ask an agent steps. In a project, the explorer on the left reads Sources → Processes → Workspace: each process is an item with a status dot (green: On; hollow: never turned on yet; amber: turned Off or paused from the tray; red: needs attention) and opens as a tab, like a file. The project's agents sit at the start of the tab bar.

---

1. Create a process

In the explorer, click + next to Processes and choose Build by hand or ✨ Create from one sentence. A new process always starts Off: it does not run until you turn it on.

Build by hand

An empty process tab opens with the Schedule row at the top. Click Schedule to set when it runs, then click + Add step to add each step (section 3). For example, an end-of-day report with a reminder:

• Schedule: every day at 17:00
  1. Fetch files · Accounting group, Drive "Reports"
  2. Check · No file from Minh today?
     • Yes → 3. Message · Accounting group: "Minh, please send the report" 4. Wait until 17:50 5. Fetch files · Accounting group
     • No → (go on)
  3. Ask an agent · Reconciliation accountant: "Summarise today's reports…"
  4. Send files · Owner — result of step 6

When you are done, review the steps and flip the switch to On (section 2).

✨ Create from one sentence

1. Choose ✨ Create from one sentence and write what should happen, the way you would tell a colleague. For the same process as above:

   At 5 pm every day fetch files from the Accounting group and Drive, if there is no file from Minh remind the group, wait until 5:50 and fetch again, then ask the reconciliation accountant to summarise the reports and send them to me.

2. AMBA asks only what it cannot tell, at most 3 short questions with answers to tap: which group, who receives it, which agent or how many files. For a group question, Connect a new Telegram group sets one up right there (see below).
3. Not quite right yet? Type a change in Refine… under AMBA's draft, as many times as you need (up to 10): "send it at 5 pm", "when it is sent, message Mr Hung". AMBA remembers your first sentence and every earlier change, runs the checks below again and answers with an updated draft that lists what changed; earlier drafts fold to one line (tap to open). Nothing is saved yet: closing the screen leaves nothing behind, and once you have refined the draft AMBA asks before discarding it. If AMBA cannot apply a change it says so, keeps the previous draft and leaves your words in the box.
4. Tap Create process: AMBA creates the process from the latest draft and opens it Off for you to review: the steps look exactly as if you had built them by hand, and you can change any of them. When they look right, flip On.

AMBA checks the draft before you turn it on

Before it hands you the draft, AMBA checks it against your sentence, so a wrong process does not break your daily work:

• See how it plays out shows what the process does in each situation — files in at the first fetch, missing and then in before the cutoff, no files all day — for example "20:00 fetch files from the Accounting group → files in → summarise the report → send the report to Chị Lan". AMBA only walks it on paper: nothing is fetched or sent. A situation in which the person who gets the report hears nothing is marked Nobody is told. You can open See how it plays out under the steps of an existing process too.
• Common mistakes show on the step: a wait and a second fetch outside the "no files yet" branch, so they run every day, even when the files are already in; a "has files" check without an Otherwise branch, so with no files, the recipients are never told; a reminder that names one time while AMBA waits until another; a fetch nothing uses; an agent that is not in the project (tap Give it to … to pick one of the project's agents).
• AMBA asks the same model once more: when a situation does not match your sentence, AMBA fixes it and shows AMBA fixed it itself: with the changed lines, or asks you one short question to tap (still at most 3). If this check cannot run (a model error), the draft is kept with a note and the situations for you to read.
• On is refused while errors are left (fix them first). With only warnings left, AMBA lists them and you tap Turn on anyway when you are sure.

Changing a process with one sentence goes through the same checks.

Ask AMBA to recheck a saved process

When a step has a warning or an error, or On was refused, the process tab shows Ask AMBA to recheck. It is also always in the tab's ⋯ menu. AMBA reads the process as it is now (the version waiting to be turned on, if any), checks each situation against what the process is for (its name, description and steps) and answers in one of three ways:

• AMBA finds the process does what it should, with any warnings left.
• AMBA suggests this fix: what does not match, the lines that would change, and the fix's warnings and situations. Apply saves it like any step edit: if the fix lets a process that is on do something new (for example message a new person), AMBA asks for your consent right there before saving. Skip keeps the process as it is.
• One short question to tap; once you choose, AMBA shows the fix as above.

Nothing is saved until you tap Apply. If the process changed after AMBA checked it, AMBA asks you to recheck. If the model fails, AMBA says why and still shows the warnings and situations. Check again next to an On error still only re-reads the process, without calling the model.

Connecting a Telegram group from a question

If no bot is connected yet, Connect a new Telegram group → Connect a Telegram bot asks for the bot token from @BotFather (the only thing you type; see Connected Accounts). Then add the bot to the group, send one message in the group and tap Reload: the group appears by its name. Turn the bot's privacy mode off in @BotFather (/setprivacy → Disable) so it can see files in the group.

You can also ask the agent in its chat ("set up a daily report for me"): its proposal shows Create process, which opens a new process tab with your words.

2. The process tab

Click a process in the explorer to open its tab. Each process has one screen. The header shows:

• its name and its state;
• the On / Off switch — the only way to turn a process on or off. Off keeps the steps and the run history. When you flip it On, AMBA shows the AMBA will: list of everything the process may do (which sources it reads, who it messages, which agents it asks); turning it on is your consent to exactly that. If the process messages a person or group AMBA has never reached, it first sends them a test message; if that message cannot be delivered, the process stays off and AMBA tells you why;
• ▶ Run whole process — runs the whole process once, now, for real. AMBA asks with one line: Run the whole process now, for real — messages and files will really be sent. Press Run to go ahead;
• ⋯ with Advanced and Delete.

Right under the header, the Next run line always says what happens next: the next run time and how long until then ("08:30 today · in 3 min", counting down each minute) with what it will do; or the run that is still open ("Running step 2/3", "Waiting until 09:00") with Cancel; or why nothing is coming up (The process is Off, No schedule yet (runs by hand only), Today's window is over — next 07:00 tomorrow).

Below are two tabs:

• Steps: the Schedule row, then the steps in order (section 3). Each step has ▶ (Run this step now, section 4); an Ask an agent step shows its agent — click it to open that agent's session for this process, in the agent's tab. Below the steps: the last result in words ("Sent to Ms Lan", "Not sent: no files today"), Test step by step and the Change it in one sentence… box (on a process with no steps yet it reads ✨ Create from one sentence…).
• Run history: the Next run card on top, then the runs by day (Today, Yesterday, day/month). Each run is one row: its time, how it started (schedule, by hand or files arrived), a status dot and the result in words ("Sent to Accounting group", "Error: …", "Cancelled"). Skips in a row fold into one muted line such as "Skipped 7 times (the previous run is still open)"; click it to see each one. Click a row to open that run's steps in place; a step opens what it took in, produced, sent or failed on, with Test this step again. Errors only keeps the failed runs. New runs appear at once, without reloading.

In the explorer, ⋯ or a right-click on a process offers Open, Run whole process, Turn on / Turn off, Rename and Delete (it asks first). A spinning mark means it is running now.

The tab says	Meaning
Not finished	A new process that has never been turned on. Review the steps and flip On.
On	It runs on its Schedule.
Paused	You flipped it Off, or chose Pause all processes in the tray (the tab then says Paused from the menu bar). Flip On to run it again.
Needs attention: …	Something stops it. The tab names it and shows one button that opens the fix.

Fixes you may see: Sign in again (Google needs you to sign in again), Connect the bot, Check the file source (the bot cannot read the group, the folder is gone), Choose (a person has no chat yet), Choose the dates (files in Sources wait for a date), See what happened (the last three runs failed), Turn back on.

Agents in the tab bar

Each agent of the project is a chip at the start of the tab bar (its initials and name; a pulsing dot while it works). Click it to talk to it; + adds an agent. Its ⋯ (or a right-click) has Agent settings, Processes it takes part in (each opens the process tab) and Past sessions (its general chat and its session for each process). Beyond five agents, the others are under Agents ▾. Above an agent's chat, Takes part in: lists its processes. Two agents you create from the same base agent are two separate agents: each has its own steps, sessions and busy dot.

The model you pick in an agent's chat is the one that agent uses in processes too: its Ask an agent steps show it with the reason ("Reconciliation accountant · gpt-6-astra (Azure OpenAI) · agent"; · project or · default when the agent has no model of its own), and the run history shows the model each step used.

3. Steps

The first row is always Schedule; the steps below it run from top to bottom. There are 7 kinds:

Step	What it does
Schedule	When it runs: Every day, Mon–Fri or Chosen days, at a time; or Repeat every N minutes (5–1440) or every N hours (1–24), within From … Until … (empty: all day) and On days you pick. Ends: Never (default), Today only, or On a date… (the last day it runs); after its last run the process turns itself Off.
Fetch files	Exactly like pressing Fetch files on a source: one, several or All sources. It fetches now and does not wait for files to arrive.
Check	An "if" branch: the steps under Yes → run when the condition holds, the steps under No → when it does not (by default (go on): carry on with the next step).
Message	Sends a Telegram message to a group or a person.
Ask an agent	A project agent does a task, with Today's files or the files an earlier Fetch files step brought in. Turn on Allow looking things up online when the task needs information from the web (gold prices, news, exchange rates): the agent can search the web as in chat; only the search queries go to the search provider you set up in Settings, and project files are not sent. Without a search key, or with Internet turned off for the project, the step fails and says where to set it up (no key: a button to Settings → Connected Accounts). If the agent itself does not have Web research, the form warns you, and a failed step shows Let the agent search the web in the run history, the step panel and the agent's conversation: one click turns Web research on for that agent, then Try again.
Send files	Sends the result of an Ask an agent step to a person or a group.
Wait until	Waits until a time (HH:MM) the same day; if that time has passed, it goes on at once. Not for a Repeat process (see section 4).

+ Add step sits between steps and inside each branch of a Check: pick the kind of step, then fill 1–3 fields with pickers (people, groups, sources and agents are chosen by name; you never type an id).

Conditions of a Check step — pick a kind; you can add several and combine them with and / or:

• No file from someone yet: "No file from Minh today". Pick the person from the project's people, or from the Telegram senders the bot has already seen (picking one adds them to the project's people).
• Files fetched this time: ≥, =, … N files (from any source or one source).
• A figure from the agent: a figure returned by an Ask an agent step, greater than, less than or equal to a value.

How "today" is counted

"No file from Minh today" counts every file of the day in the sources the Fetch files step reads, including files you fetched earlier that day with the Fetch files button. So if Minh's file already came in, no reminder goes out. AMBA recognises Minh by his Telegram account: "Mr Minh" in the people list still matches a sender whose Telegram name is "Minh Accounting". A person without a Telegram account in the list is matched by name, ignoring titles such as Mr, Mrs, Ms (and Vietnamese anh, chị, em, cô, chú, bác, ông, bà).

Changing the steps

• Click a step to open its settings in the right pane. Every change is saved when you leave the field and applies from the next run — there is no Apply button. The pane also has Up, Down and Delete step.
• If the process is on and a change lets it do more (a new source, a new recipient, a new agent, letting the agent look things up online), AMBA shows one consent line right there, for example The process will send messages to Mr Hung — Agree, before it is saved.
• Change it in one sentence… at the bottom of Steps: for example "move it to 6 pm", "also message Mr Hung", "no need on Saturdays". AMBA drafts the change for you to review first; on a process that is on, the change waits for you to press Turn on ("A change is waiting for you to turn it on."), and the earlier version keeps running meanwhile. Earlier versions are listed under ⋯ → Advanced → Version history.

Advanced steps (from earlier versions)

Processes made with earlier AMBA versions may contain steps that are not in the list above (collect reports, wait until complete, put files in order, ask back, run when files arrive). They keep running as before and show read-only as Advanced step: This step comes from an older version and still runs as before. Edit it with a sentence, or delete it and add a new step.

4. Running and testing

• ▶ Run whole process really runs the whole process once, after the confirmation line (section 2).
• A run that is still open (for example Wait until 17:50 because a file is missing) holds that day: the Next run line shows Waiting until 17:50 · Cancel, and another run that day is refused with a run for today is already open. Cancel (or Cancel this run on that run's row in Run history) asks with one line, then stops the run: its remaining steps do not run and you can run the process again right away.
• ▶ on a step (Run this step now) runs that step alone, now, using the earlier steps' results from the last run or test. The result shows in a panel right under the step:
  • on Fetch files, it really fetches, like the Fetch files button (the source's history says Fetched by process <name>);
  • on Message and Send files, you see exactly what would be sent and to whom, but nothing is sent. Send me a test sends one message labelled [Test] only to your own Telegram chat, through your bot; nobody else receives it. The first time, AMBA asks which chat is yours.
• With Repeat, each occurrence runs once. If you change the time after the process already ran today, it runs again at the new time today.
• "Message the group every 5 minutes, stop at 9": set it on Schedule — Repeat every 5 minutes, From 07:00 Until 09:00 (09:00 runs too when a repeat falls on it: 25 runs a day). Do not use Wait until or a stop step for this: in a repeating process Wait until keeps the run open, so the next repeats are skipped, and a stop step ends only that run, not the process. That is why Wait until is greyed out in + Add step of a repeating process, and such a step from before shows a warning under it. A repeat that comes while the previous run is still open is skipped (folded into one line in Run history); a schedule at a time still keeps one open run per day.
• "Every 5 minutes until 22:00, only today": set the window as above and Ends → Today only (or On a date… for "until 20/10"). The last repeat runs at 22:00 and nothing starts tomorrow: the process turns itself Off (its history stays), Run history and the Next run line say Schedule ended on 10/16, and before that the Next run line shows Last run: 22:00 today when the next run is the last. Turning it On again asks to Extend to a new date or remove the end date. Runs missed while AMBA was closed are not made up after the end date. In a sentence, "only today" or "until Friday" sets it too.

Test step by step

Test step by step in Steps turns on testing, so you can try each step before turning the process on or at any time later, without waiting for its time and without sending anything to your team.

• The panel under each step has Input (the files and the day used, and who is still missing), Result (counts, the report files — open them to preview — and the key figures), Would send (the exact message as Telegram will show it, and to whom; it is not sent) and Error (the reason and the button that fixes it).
• Test data (above the steps): Today (today's files in Sources; tick Fetch files before testing to read the sources first), Day… (any day that has files) or Previous step's result. After a successful test AMBA switches to the previous step's result by itself, so the next step runs on it with one tap.
• Simulate lets you test a branch of a Check without waiting for it: assume a source is complete or that someone is still missing (Assume complete / Assume missing), set figures (for example a loss for the day), or Assume the time is 15:00.

Tests never count as runs: they do not appear in Run history or in the process's status, and they are kept for 7 days (the latest 20). Each step of a real run in Run history opens the same panel (with Sent instead of Would send) and Test this step again.

When the model fails

If an Ask an agent step cannot run because of the model (the key is refused, the quota is used up, the provider rejects the request, the model is gone, the connection is lost), AMBA does not switch to another model. Instead it:

• Tells the people waiting for the result right away, on the process's own channel and bot: the recipients of the Send files step; without one, the recipients of the Message step; without those, the process's Telegram group. The message is short, for example ⚠️ Daily report is not done today: Google rejected the request. AMBA will try again at 22:30. With nobody to tell, AMBA shows it only on this computer and in Run history.
• Tries again twice, 30 minutes apart, in the same run (also after AMBA restarts). If it works, the report is sent as usual. When the tries run out, the people waiting get ⚠️ … has no result today: … Open AMBA to fix it (Settings → Models). Each message is sent once.
• While it waits, Next run shows Waiting to retry at 22:30 · Cancel; Cancel stops the retries and sends nothing more. The run's row in Run history says Model error: Google rejected the request · retry at 22:30; open it for Details (the provider's message), who was told and when, and a button that opens Settings → Models.
• A Repeat shorter than 30 minutes does not wait: the run stops at once and the next repeat is the retry; the same error is announced once a day. With a slower repeat, repeats that come while a retry is waiting are skipped, as when the previous run is still open.

Follow the agent live

Each time an Ask an agent step runs — on its schedule, with ▶ Run whole process, with ▶ on the step, or as a retry — it is one turn in the agent's session for this process, like a message you send in chat. The session shows the exact request the agent received (the instruction, the day, the folder for the report and the day's files) under a header such as ⏰ Daily report · Summarise the report · scheduled · 16:00, then the agent's work as it happens: its plan, each file it reads or writes, and its answer. The session keeps every run, so you can scroll back to see how the agent worked on any day.

• Watch the agent work appears on a step while it runs, in its ▶ panel and on a running row of Run history. It opens the agent's session at that run.
• While an agent works on a process, its chip in the tab bar pulses, and its own chat says Working on: …process name… · Watch.
• If the step fails, the session says why in plain words, with Details. When the agent says it created a file that is not there, you read The agent said it created bao-cao.xlsx, but the file is not there — usually the model answered without writing the file: try again, or pick another model in the agent's settings. The step's panel and Run history say the same (the agent said it created files that do not exist: …).

5. Sources and file sources

Every file that arrives is kept in the project's Sources, by source and business day; agents and processes only read it, never change it. AMBA works out which day a file belongs to from its content, title, file name, the message sent with it and its arrival time — no AI involved. A process fetches from a source only when one of its Fetch files steps runs.

In the file tree, each source is an item (for example Telegram · Accounting group) with a health dot: Receiving · No files yet · Needs sign-in again · The bot cannot read the group · The group changed address (AMBA follows the new one by itself). Each day under a source has a small badge: which report used it, files of unknown kind, files needing a date, or why the day has files but no report — click it to fix. Check the file source in a process tab opens that source's tab. Click a source to open its Ingestion history and the source's Setup (Check, Pause, Remove source). See Project Sources.

⋯ → Add source at the top of Sources adds a new source:

• Telegram — choose the bot and the group by name. The group is also saved in the project directory.
• Drive — if you are not signed in to Google, click Sign in with Google right there; then open the folder by name (My Drive, Shared drives, Shared with me) → Choose this folder. Tick Also take the files already there to take the files already in the folder; every subfolder level is read.
• Folder — pick a folder on this computer with the system picker. A folder inside the project is read in place; for a folder outside it, new files are copied into Sources and the originals stay where they are.

The new source shows in the tree at once (No files yet until the first file) and in the picker of the Fetch files step. When a process description mentions a Drive folder or a folder on this computer, the question "which file source is …?" offers Google Drive folder and Folder on this computer to add it right in the conversation. A process that mentions a group ("the accounting team") adds that group as a file source; later processes reuse it. Sources from earlier AMBA versions were moved here automatically.

Import files once… in a Google Drive source's tab copies files picked by name from Drive into that source once; ⋯ → Upload Files adds files from this computer. AMBA sorts them by day like files from a source.

6. Talking with the person who receives the report

Processes made in an earlier version from a sentence saying the owner can ask back or approve ("I can ask back and approve the bonus on Telegram") show Conversation with Ms Lan in Steps. Replies on Telegram are paused in this version, so her messages are not answered; ask the agent in AMBA instead (agent chip on the Nhờ agent step). The conversation is shown in the process tab, not as a separate agent tab.

Paused for now

Replies on Telegram are not answered yet: the conversation shows Paused (needs automatic fetching) until AMBA reads chats on its own again. To ask about a report meanwhile, click the agent of its step and ask in AMBA.

7. Keep AMBA running

Processes run only while AMBA is open and the computer is awake. In Settings → General → Background & processes, keep Keep running in the background and Keep the computer awake for scheduled processes on, and turn on Start AMBA when I sign in. Closing the window then leaves AMBA in the tray; the sidebar shows Processes on: N with the next run, and the tray menu can Pause all processes. While a process is on, AMBA asks the operating system not to slow it down in the background, so a scheduled time starts in its minute (Windows battery saver, or Efficiency mode set for AMBA in Task Manager, still wins). AMBA cannot wake a computer that is asleep or switched off; a missed run catches up within 24 hours when AMBA starts again and still works on the missed day.

8. Earlier automations

If you used automations in an earlier version, AMBA moved them into processes and file sources on the first start: the first process tab you open says how many and lists anything it could not move, with the reason. They keep their times, recipients and the bot's position in each chat, so no message is read twice. Their steps show as Advanced step (section 3).

9. Privacy

Everything a process does runs on your computer. Files stay in the project; only the text a step needs is sent to the AI model you configured, as in any chat. Messages go out only to the chats listed in AMBA will: (or that you agreed to with Agree while editing), through your own bot. Every message sent (including Send me a test) and every call to the model is listed under Data leaving this device (Project Configuration → Permissions & Security).