What is the AI Agent in Task Manager?
The AI Agent in Task Manager takes on instructions spread over time: “email a summary, and if the client hasn’t replied by Friday, let me know in the thread”. It doesn’t just give you an answer to read. It sets out steps to be carried out straight away, at a specified time or when a condition is met; a CRON job is required to run scheduled steps on time.
The difference from the Assistant is simple. The Assistant acts here and now: you ask a question and get an answer or an action carried out. The Agent handles instructions involving “later”, “if” or “keep an eye on”. A plan can have up to ten steps and expires after thirty days.
The Agent doesn’t ask for approval before starting. A plan becomes active as soon as it is created, and any steps whose due time has already passed are carried out straight away. Emails are sent from your personal SMTP, so recipients see your address. You can cancel an instruction later only in “AI Assistant” mode in a task or chat thread: start your message with “cancel plan 13” or “cancel everything”. You can’t cancel a plan in the AI Assistant window. You can only cancel your own active plans.
What does an instruction look like behind the scenes?
Task Manager first works out whether your sentence is an instruction. The question “how many of my tasks are overdue?” goes to the Assistant. A sentence that includes a time, condition or email address goes to the Agent. It sets out a plan: a short summary and a list of steps with times in the Europe/Warsaw time zone and, where relevant, conditions.
There are six types of steps. Running steps that depend on a due time - not just sending emails - requires a CRON job. Without one, these steps run with a delay and only when someone is using the app. The Agent can post a message in a thread on your behalf, send an email, notify a specific person or check the conversation history. It can also wait for someone’s reply or for a keyword in a thread. In the latter case, the step only starts when the specified person posts a message containing the agreed keyword.
When creating a plan, the Agent can only choose someone from the participant list as a step participant. When sending a message, it checks that its recipients have access. By default, messages inherit the privacy setting of your instruction: if your instruction is private, only you can see the Agent’s reply. A message explicitly marked as public is posted in the thread for everyone to see. For a public message like this, the Agent sends email notifications to the people mentioned, according to their notification settings.
What do you gain?
You don’t have to keep track of when a subcontractor is due to reply, a meeting needs confirming or a client needs to make a decision. You specify what should happen and when. If they don’t reply, you’ll be told without having to keep checking.
You also don’t have to come back in the evening to a message that needs to go out tomorrow morning. Schedule it in advance and get it off your mind. It’s worth knowing what affects whether it runs on time: without a configured cron job, steps run whenever anyone on the team is using the app. If no one is working, they wait. With a cron job, they run regardless of whether anyone is logged in.
It’s worth being careful with instructions you send just to try things out. A plan starts without confirmation, so a test instruction will work just like a considered one. Treat it like an instruction given to someone on your team, not a rough idea.
What it looks like for someone who uses it
You’ve sent a proposal and want to make sure no one forgets about it. You ask the Agent to send an email with a summary. You also ask it to take two further steps: post in the thread on Thursday morning, and take a separate step when the person responsible for the matter writes “we accept” in the thread. If the email step is already due, the email is sent from your address as part of the same request. The other steps wait for their time or condition.
On Tuesday, a thread participant specified in the plan writes “we accept” after speaking with the client. The step waiting for a keyword starts because the specified person wrote the message and used the agreed word. You receive a message in the thread.
On Thursday morning, the scheduled step starts and brings the matter back to the forefront, regardless of what happened earlier. If you’d decided on Wednesday that the matter was no longer relevant, you would have typed “cancel plan 13” in the thread.
Frequently asked questions
How is an AI Agent different from an AI Assistant?
An Assistant responds and takes action straight away, during the conversation. An AI Agent carries out instructions spread over time or dependent on a condition. A plan can have up to ten steps.
Will the Agent ask for my approval before carrying out a plan?
No. A plan starts as soon as it is created, and any steps due at that point are carried out immediately, including sending emails. The “Yes, run it” confirmation window applies to actions carried out straight away by the AI Assistant, not the Agent.
How do I cancel an instruction that has already started?
In a task or chat thread, in AI Assistant mode, type “cancel plan” followed by the plan number, or “cancel everything”. You can only cancel your own plan while it is still active. Plans cannot be cancelled in the separate AI Assistant window - that’s where you have a regular conversation.
What address does the Agent use to send emails?
The Agent only uses your personal SMTP, configured in Settings → Profile.
How long does an instruction remain active?
An instruction expires thirty days after the plan is saved.