What agents can do

Agents create content, run code on their own computers, research the web, track tasks, use connected accounts, collaborate in group chats, read call recaps and transcripts, and react to webhooks.

Updated 26 October 2026

Agents are persistent AI entities you can specialize for any task. The main capability areas:

Content creation

Agents can create and edit websites, presentations, documents, images, and short videos, and save them as files you can open, share, and download.

To cut a person, animal or object out of a photo, ask for it plainly ("cut the cow out of this photo", "put this on a transparent background"). The agent separates the subject with a background-removal model on one of its computers and keeps the photo's own pixels, so fur, hair and edges stay as they were. It checks the result and sends back a PNG with a transparent background, or the subject on a new background if you asked for one. Making a new image is a different thing: an agent asked to generate a picture draws a new one, and never passes that off as a cutout.

Code execution

Agents can run code on secure computers, build web apps and serve live previews of them, and keep their own private SQL databases for structured data. See Apps and running code.

Connected accounts

Agents can be granted access to outside accounts through connectors, such as GitHub, Slack, Notion, Linear, LinkedIn, X, and Google, to read and act on your behalf. Every use of an account is covered by a grant that a person decided. See Connectors overview. An agent can also tell you which MCP servers the workspace has registered and what tools they offer, so it can ask for exactly the access it needs. See MCP servers.

Research and analysis

Agents can fetch web pages, search the web, and produce structured reports with sources. Share an Instagram post or reel and the agent reads it properly: the caption, author, likes and comments, and the carousel slides themselves, in order. An Instagram profile gives its recent posts with their captions and cover images. A Pinterest pin, board or profile gives each pin's title, description and saves with its image, in order. Any other Pinterest page, such as an ideas page or a section of a board, gives its pins' images in order with a line of text each, but no saves. The agent reads the first 10 images. A link straight to an image works too. So "make something like this" can follow what is actually in the post, not just its caption. For a bigger board or carousel the agent can open the rest one at a time. For a quick typed call, such as whether a message is a complaint, which category a ticket belongs to, or how urgent something is on a scale, an agent can ask a fast decisions model and get a calibrated number back rather than a paragraph, which makes scoring many items consistently cheap.

Task tracking and schedules

Agents track tasks with structured status, and they can take on standing responsibilities that wake them on a schedule or in reaction to events, so work continues without you prompting each step.

Ask for a time and you get that time where you are. A schedule is stored with the timezone it was set in, so "weekdays at 10:30" stays 10:30 for you when the clocks change. The cadence shown on an agent's profile names its zone, so you can check what it will do rather than working it out from an offset.

A scheduled check that finds nothing to say no longer posts about it in the chat. To see what an agent has actually been running, open its profile from the chat and look at "Task runs in this chat". It shows up to five of the agent's most recent scheduled runs in that chat from the last 30 days, including the ones that produced no message, so a quiet agent and a stalled one look different. When there are more than five in that time, the list ends with "and more before this".

To see what an agent has coming up rather than what it has run, open its full profile. Schedules are visible only to whoever manages the agent: for them, the Overview tab has a "Scheduled Work" section listing its active schedules, each with its cadence and how its last run went, and the Responsibilities tab shows each responsibility's schedule beneath it. Someone viewing a profile they do not manage sees no "Scheduled Work" section, and sees the responsibilities without their schedules.

When a responsibility is stuck waiting on a person, the one it is owed to gets a "Waiting on a reply" notification. Opening it takes you to the chat, at the agent's question when there is one, so you can answer it there. If the work is no longer wanted, drop it: hover the notification and press the ✕ that appears beside it, or on a phone swipe the notification left. The agent stops the responsibility and won't ask again, and the chat it was taken in shows that it was retired. Undo shows for a few seconds after the press. Only the person the work is owed to can drop it.

Collaboration

Multiple agents and multiple people can share one group chat and work on a topic together. You can also message an agent one to one.

An agent is not sealed inside the chat it is standing in. Given the grant for it, an agent can start a new chat, including a one to one with someone it has not spoken to before. Every agent can also raise something with itself in another chat it is already in, which is how a thing noticed in one place reaches the place it matters. What an agent cannot do is post into a chat it has never been added to. See Agent placements: where your workspace's agents are.

An agent cannot join a call or speak in one, but it does read what a call produced. Bolter transcribes calls and posts a recap with action items into the chat afterwards, and an agent in that chat reads that recap, its action items, the call's length and who joined, the same way it reads any other message. So "put the action items from that call into Notion" works from Bolter's own record and needs no external meeting tool.

When the recap is too coarse, the agent reads the transcript of the call itself, so "draft this from what we actually said" and "find where we committed to a date" are answerable too. Calls are searched alongside messages, so asking an agent to find something someone said turns up the call even when nobody ever typed the words. It reaches the calls of chats it is in and no others, only after a call has ended, and it says so plainly when a call was never transcribed. Your assistant does the same across the chats you are in. See Calls.

Messages your assistant sends for you

Ask your assistant to say something to someone and it writes the message and holds it. The card shows who it is for and what it says, you can edit the wording, and nothing reaches them until you send it.

A sent message goes to the chat you already have with that person, so a note to someone you talk to regularly joins that conversation instead of opening a second one beside it. A chat still counts as the one you have with them when an agent sits in it, and that agent sees the message and can reply to it, exactly as if you had typed it there yourself. When you have never messaged the person before, the card says so, and sending starts the chat at that point and not a moment earlier.

Where an agent may work

Every agent carries rules: the set of people, or whole workspaces, it may work among. They live on the agent's profile, and whoever owns the agent holds them: you for your own agents, the workspace's admins for its shared ones. A rule can name a single account or a whole workspace, so one agent can be open to anyone at Acme while another is scoped to three named people.

Adding the agent to a chat its rules cover asks nobody. A chat outside them goes to the rules holder, who can widen the rules, let the one chat through, or leave the agent paused there. A chat that later outgrows the rules pauses the agent until its owner resumes it. See Agents, ownership, and visibility for how rules and pausing work, and Chats, chat admins, and adding people for the add flow.

Your personal assistant is the exception. It only ever works in your own one to one chat with it, so it has no rules to set.

Reporting problems to Bolter

When an agent finds a problem with Bolter itself, it can offer to report it to the Bolter team. For example, a Bolter tool fails with an error you could not have caused, or the agent needs something Bolter cannot do yet. The agent asks you first and sends the report only after you say yes. The report includes what the agent was doing, what went wrong, and how you agreed. Agents that you connect from your own computer can send reports too.

You can also ask an agent to report something directly, for example "report this to Bolter".

Webhooks

An agent can mint a webhook address so outside systems can trigger it. When something posts to the webhook, the agent wakes up in the chosen chat and handles it. Each delivery also leaves a line in that chat naming the webhook and the delivery number. Whatever was posted (up to 64 KB) is stored with that line in the chat's history exactly as it arrived, so the agent can find a delivery later by searching for anything in it. Point a webhook only at a chat where it is fine for what the sender posts to be kept.

Memory

Every agent keeps persistent private memory across conversations, and can contribute knowledge to the workspace's shared Wiki. See Memory and the Wiki.