The Customer Service Agent Preset

Support is where wrong answers cost the most, so this preset is built around one structural rule rather than around model judgment: no verified knowledge, no answer. That single constraint changes the character of automated support entirely, every reply your customers receive traces back to something you wrote or approved, and the system's coverage grows in exactly the order your customers ask for it.

What It Handles

Three intake channels feed the agent. Email arrives through the IMAP account you configure, and replies go out through SMTP, your own support address, with the conversation threading customers expect. Web forms from your site land in the same intake. Chatbot escalations arrive when the chatbot meets a support question in any connected channel and hands it to the specialist. Whatever the door, the handling is identical: identify the question, search the knowledge base, and reply only on a verified match.

The Verified-Answer Rule

When a question arrives, the agent searches the knowledge base by meaning, so phrasing never blocks a match, a customer asking "can I get my money back" finds your refund policy. On a confident match, the agent composes a reply from the verified content, in your brand voice, addressed to the customer's actual situation. Without one, it does something more valuable than improvising: it raises a flag.

This is the property that makes the preset safe to run on real customers from day one. The model writes fluent prose, but the facts in that prose come from your knowledge base or the reply does not happen.

The Flag Loop: Answer Once

Unanswerable questions arrive in your flags panel with full context. You answer, in the UI or by telling the master agent, and two things happen: the customer gets their reply, and the answer joins the knowledge base permanently. The next customer who asks gets it automatically.

The arithmetic works strongly in your favor. Customer questions follow a steep curve, a handful of questions make up most of the volume, so every answer you bank removes a recurring interruption. New installs typically see the flag rate fall week over week as coverage builds, and what remains flagged is genuinely novel, exactly the questions that deserve a human.

What Counts as Verified

Verified knowledge accumulates from three directions, and all three end in the same searchable store. Imported documents are the foundation: your FAQ, policies, manuals, and pricing, loaded on day one. Flag answers are the growth engine: every question you settle becomes a permanent entry, written in the moment it was needed. And the research pipeline is the bulk builder: point it at a topic your customers keep raising, and its fact-checked output can deposit straight into the support knowledge base. Three sources, one rule, if it is in the base, you put it there or approved its path in, which is what lets you stand behind every automated reply with a straight face.

Setting It Up

The agent needs IMAP and SMTP credentials in config/customerservice.json, edited with validation in the Config tab's Customer Service panel, and it is gated until they exist, an enabled-but-unconfigured agent skips its AI call and costs nothing. Two preparation steps multiply its usefulness on day one: import what you already have into the knowledge base, your FAQ, policy pages, and best support replies, and fill in the support policies file in memory/system so its tone and boundaries match your standards. The master agent's onboarding covers all of it conversationally.

Running It

The standard controls apply: its own tick, model, and pause in the Models panel, Kill Now for the current task, and the global pause. The supporting tool keeps customers, replies, and resolution state inspectable from the command line whenever you want to see the queue directly. Model choice here rewards a mid-tier pick, the facts come from the knowledge base either way, and a better model writes warmer, more precise replies around them.

The operating habit that matters is answering flags promptly, each one is a customer waiting and a permanent knowledge gain from a minute of your time. Owners who clear flags daily reach the quiet state fastest: an inbox that answers itself, correctly, in your voice, around the clock.