Getting your AI Agent set up is the start, not the finish. The agents that perform well over time are the ones that get reviewed, refined, and kept current. These practices apply whether you've recently launched or you've been running the agent for months.
Get the personality right before anything else
The quality of the agent's replies depends more on your personality setup than your knowledge base. A well-defined personality with good message examples sounds on-brand even when the knowledge base has gaps. A thin personality with great knowledge still sounds generic.
Before you go live, add at least five message examples from real conversations: a warm welcome, a pricing question, a complaint, a follow-up, and a time you had to say no. Those examples shape the agent's voice more than any written rule. See Configuring Your AI Agent's Personality and Tone for the full setup.
Build the knowledge base around your FAQs
Start with the questions you answer by hand most often. Open your last month of DMs, find the ten questions that keep coming up, and write clear answers to each. Add them as manual text entries.
That covers most conversations from day one. Add more detail later as you spot gaps. Write each entry the way you'd answer the question out loud, not as a formal document. Conversational source material produces conversational replies.
Test before you go live
After setup, use the Chat with Agent button on the AI Agent page to run a real conversation. Ask the things a new follower, a confused customer, and a difficult contact might ask.
Look for three things: is the answer accurate, does it sound like you, and is the length right for the question? If any are off, fix them in the personality or knowledge settings before you enable the agent.
Start narrow with activation rules
If you're unsure how the agent will do with your audience, start in Specific intents only rather than All pending messages. Define two or three clear intents and let the agent handle only those conversations first.
That builds your confidence in the output before you open it up to every DM. Broaden the rules once you've seen how it handles real conversations. See Activation Rules.
Review conversations in the first week
For the first week, check the agent's conversations daily. Open the inbox, filter for conversations the AI Agent is handling, and read through them.
You're looking for wrong answers, off-brand replies, and moments the agent couldn't answer but should have. Each gap points to a fix: a missing knowledge entry, a style rule to add, or a message example to write.
Keep the knowledge base current
The agent can only answer accurately about what you've told it. When your pricing changes, a policy updates, or you launch something new, update the knowledge base too.
Set a monthly reminder to review your entries. Check that they still match what you actually offer, and remove anything out of date.
Know when to step in
Not every conversation suits the agent. High-stakes negotiations, complex complaints, and anyone who's asked for a person should go to a real person. Turn on Human intervention so the agent hands those off on its own, and take over manually from the inbox whenever the situation calls for it. See Human Intervention: Handing Off to a Human.
Advanced: model, reasoning, and your own OpenAI key
Most accounts never need to touch these, but they're worth knowing about.
Model. By default the agent uses a fast, low-cost OpenAI model (
gpt-5-nano) for replies. You can point it at a more capable model for higher-quality answers, at more cost and a little more latency.Reasoning. This sets how hard the model "thinks" before replying, from
noneup throughlow(the default),medium,high, andxhigh. Higher levels handle nuanced, multi-step messages better but are slower and cost more.
The model and reasoning settings aren't in the standard settings screens. They're advanced options you set through the Inrō API or MCP server, or by asking support. Leave them at their defaults unless you have a specific reason to change them.
Bring your own OpenAI key. You can add your own OpenAI API key from your billing settings. AI calls then run on your key, and Inrō checks the key is valid when you save it. Using your own key also enables the option to include a knowledge item in every reply (rather than only when it's relevant), described in Knowledge Base: Teaching Your AI Agent What to Know.
🐾 Netsuke's Tips
If the agent keeps struggling with one kind of question, add a message example showing exactly how you'd handle it. A single good example often fixes a persistent issue that style rules couldn't.
Keep knowledge entries focused. One entry per topic beats one long document covering everything. The agent finds and uses tightly scoped entries more reliably.
Review your intent-based actions every month. Check which intents fire often and which never do. Low-traffic intents usually need rewording to match how your audience actually writes.
What's next?
The agent gets more useful when it works alongside your scenarios. See AI Smart Actions in Scenarios to hand off to the agent mid-flow or branch your automations with AI.