Choose what an agent may do on its own
Set a rule for each kind of change an agent can make (go ahead, ask me or never), cap what it spends per day and approve its changes in its chat.
An agent can always look things up. What it may change, send, post or delete without you is your choice, one rule at a time.
For your main agent, these rules apply to its routines, which run while you are away. When you chat with it, it does what you ask, and anything that sends, posts, deletes or spends an app's credits waits on a card in its chat for your approval.
Pick what it may do
- Open Agents and pick the agent.
- In the panel beside its chat, open Details, then choose What Maya may do (with your agent's name in place of Maya; a short summary of its rules shows under it).
- Under Before Maya..., each kind of change has its own rule:
- Adds or edits records
- Sends a message or email
- Posts publicly
- Deletes anything
- For each one, pick Go ahead (it does it on its own), Ask me (it waits for your OK in its chat) or Never.
Each choice is saved as soon as you make it. Use Go ahead only for an agent you trust. A new agent starts with Ask me for records and messages, and Never for posting publicly and deleting.
The line under the chat box always says what the agent may do, for example "Maya asks before changes and messages, and never posts or deletes."
See which apps it can see
At the top of the same screen, Maya can see lists where it works on StartupCorners (Your StartupCorners account for an agent that works just for you) and each connected app. Untick an app to keep this agent away from it. Choose + Connect another app to go to the Apps tab.
Shared apps connect to companies: an agent that works for just you, a community or a program uses no shared apps for now. Any agent can still post in your own Slack, Discord or Telegram channel (below).
Let it post in your Slack, Discord or Telegram channel
- Connect Slack, Discord or Telegram on Apps, and in its panel pick the Channel for agents under Agents can post here. See Connect Slack, Connect Discord or Connect Telegram.
- Open the agent's Apps tab. Under Posts to, turn on the switch next to that app. Or turn the agent on under Agents that may post in the app's panel.
A community's agents can also post in a channel an admin of that community shared with it, under Share with a community you run in the app's panel. It shows under Posts to with "shared by" and that admin's name. An agent whose owner picked a channel of their own for that app keeps posting there.
A post follows the Sends a message or email rule. With Ask me, its card in the chat shows the channel and the whole text, for example "Post in #updates on Slack", before anything goes out. Every post ends with the agent's name, and an agent posts at most 3 times in one chat turn or routine run.
An agent switched on for Slack or Discord can also read the last messages of its own channel. It only posts, and never reads, in a channel another admin shared with its community, and your main agent never reads a channel. Telegram does not let bots read a channel, so there agents only post.
A post waiting for your OK expires when you turn the agent off for that app, pick another channel, or disconnect the app. On Telegram it also expires when your bot leaves the channel or stops being an admin there. A community agent's post waiting for a shared channel expires when that admin stops sharing it, picks another channel or disconnects the app.
Set a daily limit
- On the same screen, find Credits per day, at most.
- Type a number of credits from 1,000 to 500,000. Leave it empty for no limit.
- Press Enter or click away to save.
The line under it shows how many credits the agent has used today. They come out of your AI credits (see How AI credits work). The day follows your own time zone.
Once the limit is reached, until the next day:
- A message you send gets the answer "Maya has reached today's limit of 5,000 credits. Raise it in What Maya may do, or try again tomorrow."
- Run now on a routine is refused with the same words.
- A scheduled routine is skipped, and a line in the chat says so. The routine stays on.
- A routine that reaches the limit while it runs stops, and its post says it stopped before finishing.
Raise the limit here to let the agent carry on today.
Approve or reject its changes
When the agent wants to change something, a card in its chat lists the changes, for example "Maya wants to make 2 changes".
- Choose the tick to approve a change, or the cross to reject it.
- With several changes, Approve all or Reject all decides them together.
You can also just write "go" (or "yes", "ok", "do it") in the chat: it approves every change on the agent's latest card, like Approve all, and the agent carries on. A longer message, like "go but skip Naboo", is a message, not an approval. A request for more freedom always needs its own tick.
Each change then shows Done, Rejected, Failed or Expired. When a connected app turns a change down, the card gives the app's own reason, for example "Clay said: Missing required field: Company Domain". The number of changes waiting for you shows beside Agents at the bottom of the menu, and beside the agent in your list.
A change in a connected app says what it does in plain words, for example "Run Work Email in Clay for 4 people", with a small table of what it sends (each person's name, company and domain), so you can spot a wrong match before the app spends its credits. Long lists show the first 5 and how many more.
Changes in your CRM read the same way, for example "Update Diliara Lupenko in the CRM: Emails set to dlupenko@smile2impress.com".
Once you approve, the agent carries on in its chat: a line like "Approved: Run Work Email in Clay for 4 people" shows, then the agent fetches what an app sent back, moves on to the next step (like the next batch), or says in a sentence what was done. Anything it then wants to change, like saving those emails in your CRM, waits for you again. That reply uses AI credits like any other.
When an app takes a while, like Clay finding emails for 40 people, the agent checks again on its own instead of asking you to write later: a line like "Checks again at 1:44 PM" shows in its chat, and when the results are in it posts them there and you get a notification. If the app is still working, it checks again a little later (a dozen times at most).
Change one rule in the chat
You can also tell the agent in its chat, for example "never post publicly" or "go ahead with changes without asking". Less freedom applies at once. More freedom shows up as a change marked More freedom, which you approve on its own: it is never part of Approve all.
Outside content always waits
Once an agent has read something written by people outside your team, like a connected app, your Slack or Discord channel, the feed, a profile or a join request, anything it would change for the rest of that run waits for your OK, and it does not change its own notes, routines or rules, even when a rule says Go ahead.
