How to manage your agent

Where you talk to it

You can reach your agent on Telegram, on WeChat, in the web chat in your Command Center and in the desktop app, depending on what we've set up for you. It's the same agent everywhere, with the same memory, wiki and skills.

Each chat has its own conversation. In a group chat, every person gets their own conversation with the agent, so two colleagues don't trip over each other's requests.

Memory: what it knows about you by heart

Your agent keeps two small notebooks. One holds notes about your business and how you like work done, about 2,200 characters. The other is a profile of you, about 1,375 characters. Both are loaded at the start of every conversation, which is why it doesn't need reminding who you are.

They're small on purpose. Memory is for stable facts and preferences, and anything longer belongs in the wiki.

How things get saved

• It saves on its own when something is worth keeping, and it reviews the conversation every few turns for things it missed. You'll see a short "Memory updated" note when that happens.
• To be sure, say it outright, for example "Remember that I want reports as PDF on Monday mornings."
• To correct something, say "That's outdated, we use QuickBooks now, not Xero." It replaces the old note rather than adding a second one.
• To remove something, say "Forget the note about the old office address."
• To check, ask "What do you have in memory about me?"

A new memory is used from your next conversation. In the conversation where you said it, the agent already has it, because it's right there in the chat. When the notebooks get full, the agent merges and trims older notes before it adds new ones.

What goes where

The wiki: your agent's reference library

We set up a wiki in your agent's workspace, a folder of linked notes with an index and a log of changes. Your agent is told to check it first and to keep it up to date. It's where documents, procedures and anything too long for memory should go.

Adding to it

• Send a file in chat and say "add this to the wiki".
• Paste or dictate the text and say where it belongs, for example "add this to the wiki under refunds".
• Send a link and ask it to read the page and save what matters.

Plain text and web pages work everywhere. Some document formats need a converter we switch on for you, so if your agent says it can't open a file, tell us.

Keeping it current

• When something changes, tell the agent, for example "the price list changed, here's the new one, replace the old page".
• Ask what it has before you rely on it, for example "what does the wiki say about deposits?"
• Ask it to mark old documents as replaced rather than delete them, so the history stays readable.

Skills: routines it does the same way every time

A skill is a written procedure, the steps, the checks and sometimes a small script, that your agent loads only when a task needs it. Your agent comes with a set of general skills, we add skills built for your work, and it writes its own.

• After a task that took several steps, it may save the method as a skill on its own.
• You can ask for one directly, for example "turn what we just did into a skill called month-end-close".
• Every skill can be started as a command, so later you can type /month-end-close.
• To see what it has, ask "which skills do you have?"
• To change one, tell it what to fix, for example "in the month-end skill, also export the aged debtors list". It edits the skill rather than starting again.

On the current version, skills the agent made itself and hasn't used for 90 days are archived, not deleted, and we can bring them back. Skills we build for you are kept.

Conversations

Each chat is one ongoing conversation. It doesn't reset at midnight, and it carries on until you start a new one.

Long conversations

When a conversation gets long, your agent summarises the older part automatically and keeps the most recent messages word for word. That keeps it fast, but fine details from far back can get lost in the summary, so put anything that must last into memory or the wiki.

Past conversations are searchable

Every conversation is stored, and your agent can search them. Ask "what did we decide about the Tanaka quote last week?" and it looks it up rather than guessing. Naming the work helps a lot, because "the Q2 marketing plan" is easier to find than "that thing we discussed".

When to start fresh

Type /new when you switch to an unrelated task or when replies start to drift. Type /resume to go back to an earlier conversation, and /title to give the current one a name you'll recognise. Finished conversations that sit unused for about 90 days are cleared, so anything you'll need later should be in memory or the wiki.

Scheduled jobs

Ask in plain language and your agent sets the job up.

• Results come back to the chat where you asked, unless you tell it to send them somewhere else.
• To manage jobs, ask "list my scheduled jobs", "pause the Monday report", "run it now" or "delete the reminder".
• Each run starts fresh, without the chat around it, so put everything the job needs in the request, or point it at a skill.

If we move your agent to a different model, we re-check its scheduled jobs, because the safety setting pauses jobs when the model under them changes.

Files, images and voice

• Images. Send a photo, a screenshot or a scan and it reads it, using a model that can see images.
• Voice notes. They're transcribed on your agent's own server and answered like text. If you speak a language other than English in your voice notes, tell us and we'll set the language.
• Voice replies. Type /voice on and it answers your voice notes with voice.
• Documents. Text files and web pages work everywhere. If a format won't open, tell us.

Approvals and safety

• Some actions wait for your OK. We agree with you during the build which ones, for example anything sent to a customer or anything involving money. Reply yes or no, or use /approve and /deny.
• A hard block on destructive commands always applies, whatever the settings.
• Never paste passwords, card numbers or API keys into the chat. We connect accounts for you outside the conversation.
• In your browser, the agent stops and hands control back for logins, captchas and payments.

Commands worth knowing

For skills, scheduled jobs and settings, just ask in plain words. Commands like /skills and /cron belong to the command-line version and don't work in chat.

When something seems off

For anything else, email [email protected].