Conversations, clarifying questions, @-mentions, model choice, and what happens after it acts.
The Agent is a real conversation, not a fixed script — you can type anything, and it decides what to do with it: search, report, draft, or ask you something back first. This page covers the mechanics that make that work.
A model selector sits right next to the send button: gpt-6-luna (free, quickest), gpt-6-sol or claude-opus-5-5. On gpt-6-sol and claude-opus-5-5, an effort picker next to it (Low, Medium, High) sets how deeply the agent reasons on each turn, and each level shows its estimated credit cost. gpt-6-luna always runs at Low.
The "+" menu's "Attach a file" accepts anything — but a CSV or Excel file is auto-detected as a lead list and handled specially (see Importing leads from a file or CRM).
The same "+" menu has a CRM option — connect HubSpot, Pipedrive, or Zoho right there, or import from one you've already connected, without leaving the chat.
Press the Up arrow in an empty composer to bring back your last message in this conversation, shell-history style — Down arrow moves forward again.
Replies stream in live — you're not staring at a blank spinner while the Agent thinks. As it works, you'll see it happening in real time, one step at a time:
Each lookup or action the Agent takes shows up as its own line — a pulsing icon while it runs, a checkmark the instant it finishes — so you can see it's searching, checking performance, or saving leads, not just "thinking."
Once the Agent starts answering, the text streams in word by word, the same conversation you'd get from typing it yourself — not a wait-then-reveal-everything-at-once pause.
Two kinds of answers get their own purpose-built card instead of a plain paragraph — everything else still renders as clean markdown (headings, lists, bold, tables).
Renders as a card with Subject and Body clearly separated. A multi-email sequence gets a tab strip — Initial email, Follow-up 1, Follow-up 2 — and each one has its own independent Edit, Copy, and Save as template buttons, so you can save just one step of a sequence without the rest.
Renders as KPI cards (Sent, Open Rate, Reply Rate, and more) up top, a side-by-side comparison table when you asked about more than one campaign, and a separate "AI insight" panel for the Agent's own read on the numbers.
Expect a tabbed card, one tab per email — save any single step as a template on its own.
Expect KPI cards plus a comparison table — rates are correctly weighted across campaigns of different sizes, never just averaged.
Type @ to open a picker over four categories — Campaigns, Leads (your saved lists), Templates, and Email Accounts (senders) — and reference something real by name instead of describing it. The mention highlights inline as you type, and once sent, becomes a clickable link back to that item.
When your request is missing something important, the Agent asks — as an interactive card, not more prose to parse.
| Field | What it means |
|---|---|
| A single question | Clickable options (with a "Recommended" pick flagged) or an editable suggested value — answer with one click, or type your own via "Something else…". |
| Several questions at once | A short paged wizard docked above the composer, one question per screen, each pre-filled with a sensible default. Press 1–9 to pick an option, or Skip to move on without answering. |
Checking performance, replies, credits, or a template's content doesn't change anything — the Agent just reports back, no confirmation needed.
Building a campaign, launching a re-engagement send, or starting a campaign always renders a review card or step-by-step flow under the reply — you click the actual button (Start Campaign, Launch, Save) to proceed.
A card that just created or started something links straight to that campaign, template, or list — so you can pick up where the chat left off.
Next up: how the Agent actually finds companies and people when you ask it to prospect.