AI Agent 101

The AI Agent is a chat assistant built into Topo. It sits next to whatever you are working on and can do almost everything you can do yourself: run a lead search, build a sequence, pull your analytics, suggest ideas, create a playbook. You talk to it, it works on the page in front of you, and you watch the result appear live.

You will find AI Agent as the first item in the left sidebar.

Opening the agent

Click AI Agent in the sidebar. A side panel opens on the right with the chat, and the rest of the screen stays visible behind it. This is the default view, and it is the one to use when you want to see the agent's work as it happens.

You can also open it with Cmd+J from anywhere in Topo, or via the sparkles icon in the top bar. Cmd+K offers an Ask the AI agent option that carries over whatever you have already typed there.

The icons at the top right of the panel let you:

In full screen you get the chat on its own, with no page behind it. Useful when you are planning or asking questions rather than watching something being built.

What the agent already sees

The agent knows the page you are on, the record open on your screen, any filters you have applied, and the rows you have ticked. That is why asking to "enroll these" on a list with 40 rows selected enrolls exactly those 40, without you needing to re-describe what you are looking at.

Suggested actions

A new chat opens with four suggested actions:

• Get analytics for a read on your current outbound performance and stats

• New search to start a new lead search straight away

• Suggest an idea for list and sequence ideas, based on what the agent knows about your organisation

• Generate a new sequence to kick off sequence generation

These are shortcuts, not limits. You can ask the agent for anything in plain language.

Giving the agent context

Two things help the agent get it right first time.

Attachments. Use the paperclip icon in the message box to attach a file.

Mentions. Type @ in the message box to point the agent at something that already exists in your workspace:

• Skills

• Lists

• Company lists

• Sequences

• Lead searches

• Individual leads

• Playbooks

So instead of describing a sequence, mention it and the agent reads it directly. Mentioning a skill is also how you force the agent to use that skill, rather than leaving it to decide. 📄 Skills 101.

Anything the agent names back to you, a lead, company, list, sequence, lead search, task or playbook, comes out as a clickable chip: click to open it, hover to preview key details. It works the other way too: paste a record's link into the chat, or type @ and pick it, and it becomes a chip.

Mentioning a record is talking about it, not instructing the agent to change it. The agent will read a lead you mention, but it will not edit, enroll or delete it unless you actually ask.

Write to the agent in whichever language you work in. Write in French and it answers in French.

How the agent works through a request

The agent shows its thinking rather than hiding it. In a typical request you will see:

1. Reasoning steps, each with a short line explaining what it is doing (for example Clarifying Search Scope, Refining User Queries).

2. Clarifying questions where information is missing. These appear as a short guided form: Step 2 of 3, a question such as "Which geographic region should we target?", and a set of options to pick from, including Something else… if none of them fit. Use Previous and Next to move through the steps, then submit.

3. The plan it is about to run.

4. The work itself, executed live on the page.

The clarifying questions are the point where you steer the outcome. Answering them is faster than correcting the agent afterwards.

Example: building a lead search

Ask for a new lead search and the agent will:

1. Load its Lead search skill.

2. Ask which target audience segment to find leads for (for example Sales & GTM leaders), which geography (for example Western Europe), and what volume to aim for (for example 1,000 leads).

3. Define the ideal audience and show you the plan.

4. Build the search live. On the left you see the Lead Search page filling in: job titles, locations, excluded job titles, company size, industry. In the chat you see each decision as it is applied, for example Excluding junior roles, Excluding CTO roles from org context.

5. Report back with a Search Overview (total leads found against your target, target audience, company profile, geography) and a Sample Results Preview table of five to six leads with name, title, company and location. This means you can follow what happened even in full-screen chat, without looking at the page.

6. Recommend next steps, usually Refine titles or Save & Import into a contact list.

Ask it to change the search ("add Business Development Director and Manager") and it updates the filters and the results in place, then asks again whether you want to refine further or save the list.

Beyond lead search and sequences

The lead search walkthrough above is one example. In practice, anything you can reach from the left sidebar is fair game: import leads, manage exclusions, create and assign tasks, categorize a reply, enrich a person or company, update a CRM field, change a sequence's status, senders or settings, run a playbook, or research a company from the open web. On the page you are on, it can also navigate somewhere, open a record, change what a table is showing, or select rows for you.

It can also answer questions about how Topo works by searching the product documentation, and it can search, read, create and reply to support tickets.

Where you can use the agent

The agent is not only in the app.

• In Slack. Mention @Topo in a channel, or message the Topo app directly. Same agent, Slack-formatted replies. See 📄 Using the AI Agent in Slack.

• On a schedule. Ask for a job to run once at a set time or on a repeat. The agent runs it in the background and sends you the result. See 📄 Running the AI Agent on a schedule.

• Against your other tools. Connectors let the agent read your call recorder or your knowledge base, and you can add any custom MCP server. See 📄 Connectors.

Wherever you launch it from, it is the same agent with the same access. That includes your CRM, so you can ask it things like which deals closed last month, as well as your analytics and your sequences.

The agent asks before it acts

The agent does not act behind your back. It proposes what it is about to do and waits for you to confirm. Nothing is created in your workspace, and no lead is contacted, until you have said yes.

Three things worth knowing:

• You can stop the agent mid-run. If it is heading the wrong way, interrupt it.

• You cannot undo what it has already done. The agent works under exactly the same limits you do: if an action cannot be reversed by you in the app, un-enriching a list for example, the agent cannot reverse it either.

• Talking to the agent does not consume credits. Only enriching leads does, exactly as it would if you did it yourself.

What the agent can and cannot do

As a rule of thumb: anything you can reach from the left sidebar, the agent can act on and answer questions about.

Two settings shape how well it performs, and both are worth setting up:

• Agent Memory stores durable facts about your organisation, so you do not repeat yourself. See 📄 Understanding Agent Memory.

• Skills tell the agent how to handle a specific type of request. See 📄 Skills 101.

Every Topo user has access to the AI Agent. Your chat history is personal and is not shared with the rest of your team.

Your conversations

The history icon lists everything you have asked, searchable and grouped by Today, Yesterday, Previous 7 days, Previous 30 days and Older, with titles written automatically.

Coming back to the agent after a break of more than eight hours starts a fresh conversation; the previous one stays in your history. In a long back-and-forth, earlier turns get condensed into a summary so the agent keeps the thread, though fine details can fade. Attaching a file to a message means the agent reads it in that exchange, but reopening the conversation later will not put the file back in front of it.

The agent is scoped to prospecting, outbound, go-to-market and support, and declines anything outside that on purpose. A single request also has a time ceiling: a very heavy multi-step ask can hit it, so split the work into two if that happens. The stop button cancels on purpose too; a cancelled card reads "Cancelled".

Still have a question? See 📄 FAQ: AI Agent, Memory, Skills & Playbooks.