An Agent represents you as a seller. It holds your company context, your product positioning, your tone, and one or more Personas (ICPs you're targeting). Every campaign you run is attached to an Agent, and the Agent is what tells Twain who is doing the outreach and why someone should care.

Agents serve two purposes: research and qualification (Signals tell Twain what buying indicators to look for, Warnings define what disqualifies a lead), and continuous improvement. As you work with campaigns and refine what's working, you can push those changes back to the source Agent so every future campaign starts from a better baseline.

Click Create Agent and enter your company website domain. Twain reads your site and pre-fills the setup with your product description, positioning, and suggested Personas. Review and refine what it generates before moving on.

During setup you can also provide additional collateral: upload a file (an lm.txt, a pitch deck, a one-pager) and Twain will factor it into the Agent configuration alongside your website.

When you first create an Agent, Twain asks you for a setup prompt. This is a one-time briefing: describe your purpose and audience, add any research instructions (signals you want Twain to look for) and disqualifiers (who you don't want to reach).

Twain reads this and automatically sorts everything into the right fields, Description, Signals, Warnings, and Personas.

The Description field itself is just Purpose and Audience: what your product does and who you're selling it to. What stays true across every campaign you run with this Agent.

What you're doing in this specific campaign (the occasion, the angle, the offer) belongs in the brief, not here.

After you complete the setup prompt, Twain automatically proposes buyer Personas based on what you described. Review them, adjust what doesn't fit, and add any segments Twain missed.

Personas live inside the Agent under the Personas tab. Each one defines a target role, their pain points, objections, social proof, and competitive advantage. You can have multiple Personas per Agent, which is useful when you're targeting different roles with the same product and want the outreach to reflect those differences.

During Agent setup, Twain automatically generates Signals and Warnings based on your setup prompt and shows them to you in the Rules step for review. You can edit, add, or remove anything before you create the Agent.

Signals and Warnings are the two sides of lead qualification.

✅ Signals ("Data to look for") tell Twain what to actively look for during research, on top of what it already finds by default. Use them when you want Twain to prioritize specific insights, search criteria, or triggers. For example: "Check if they've posted about outbound sales on LinkedIn" or "Look for companies that have recently hired an SDR."

⭕️ Warnings ("Qualification rules") tell Twain what disqualifies a lead. These become the rules behind the warning system: any lead that matches a warning rule gets flagged. For example: "Flag leads who are not in a decision-making role" or "Flag if the company has fewer than 10 employees."

Twain already researches key lead data and flags common issues by default. Signals and Warnings let you add your own criteria on top of that. You can update them anytime from the General tab in your Agent settings.

Inside the Agent, the Writing Style settings let you:

Yes. The Agent chat on the right side of the editor is context-aware and has web access, meaning it knows exactly what's in your current setup and can look things up to supplement it. Click Review my setup and it will analyze your configuration and suggest improvements. You can ask it to apply changes directly from that chat.

You can also use the chat to brainstorm: ask it to sharpen a Persona, suggest new Signals, or stress-test your positioning against a specific competitor. Think of it as a collaborator that already knows your full setup.

Click + in the sidebar and select Campaign. The setup has three steps: Leads, Agent, and Context.

Leads — upload a CSV or connect directly from your CRM or sequencer. LinkedIn URL is the most important field; it's what Twain uses for personalization. Email is needed for export.

Agent — pick your Agent and choose which Personas to include. You don't have to activate all of them for every campaign. If you're targeting only one role in this specific campaign, toggle the others off.

Context — this is a guided chat. Twain asks you a series of questions to understand why you're reaching out right now: the occasion, what you want the lead to do, any relevant context. Focus on your intent, not on research rules or qualification criteria. Those are already stored in your Agent. Once Twain has enough context, the Generate button unlocks.

One thing worth knowing: because the Agent already handles research and qualification, the Context step is lighter than it looks. You're not re-explaining your product or your ICP. You're just telling Twain what's specific about this campaign.

Versions let you save a snapshot of your Agent's current setup before making changes. Click Versions in the top right of the Agent view to see your saved versions and create new ones.

The use case: you have a setup that's working well, you want to test something different (new signals, different warnings, adjusted positioning), and you want a fallback in case the changes don't work out. Save a version first, make your changes, and you'll always know what the previous state looked like.

Current limitation: Restoring and deleting versions is not yet possible. You can create new versions and adjust existing ones, but you can't roll back automatically. If you need to return to a previous state, use the saved version as a reference and manually revert the fields.

When you adjust your agent during campaign setup or in the campaign chat — changing a persona, updating a warning, tweaking tone — those changes are saved locally to that campaign only. They don't automatically update your source agent.

Twain tracks all of these as local changes and surfaces them in the Updates tab of your agent view inside the campaign. Open any campaign, click the gear icon next to the agent name, and go to the Updates tab. You'll see two counters:

Each local change is listed by what was modified (for example, "Updated: Retail buyer — Changes in Solutions, Objections, Social proof"). The right panel shows a diff: strikethrough text is what was removed, normal text is what replaced it.

It depends on whether the change is an improvement or an experiment.

Publish if: the change makes the agent genuinely better and you want it to apply to all future campaigns. For example, you refined a persona's pain points based on real customer feedback, or you added a warning rule that's consistently catching bad leads.

Keep it local if: you're testing something new and haven't seen results yet. Run the campaign, check reply rates, then decide. Publishing before you have data means you might bake in a change that doesn't actually work.

To publish, click Publish to Source Agent at the bottom of the Updates tab. This pushes all local changes at once. To publish individual changes only, use the Publish button next to each item.

No. One Agent can power multiple campaigns. You'd typically create a new Agent only if you're selling a different product or representing a different company.