Every agent in eChat has a handful of settings that shape how it sounds, what it knows, and how it works with your human moderators. For agents linked to eWebinar — which is every agent the setup wizard creates — these live under three screens in the agent editor: Instructions, Product, and Agent.
None of these are set-once-and-forget. As your product changes, your training material grows, or you notice the agent handling something oddly, you come back to these screens to adjust it. This guide walks through what each one controls and how to use it well.
Instructions
What it is
The Instructions field is a plain-text box where you tell the agent how to behave: its tone, what it should focus on, and what it should never do. It is the closest thing to giving the agent a personality and a set of rules.
What it's best for
- Setting a tone - casual and friendly, or formal and precise
- Steering the agent toward certain topics, like pricing or onboarding steps, and away from others, like competitor comparisons or refund exceptions
- Hard rules, such as never quoting a price without a disclaimer, or always recommending a demo before checkout
- Handling edge cases you've noticed coming up in real conversations
How it works
When you first set up an agent, the setup wizard asks you to pick a Role - Sales, Customer Success, or Support (the default) - and seeds the Instructions field with a starting set of behavioral instructions matched to that role. If you switch roles after editing the instructions, eChat asks before replacing your edits. That's just a starting point. You can rewrite any part of it at any time from the agent's editor. There is no fixed format - it reads like a note to a new hire, not code. The agent treats these instructions as standing guidance for every reply, on top of whatever it pulls from its training material.
Best practices
- Write specific rules, not vague ones - "never mention pricing under $499" works, "be helpful" doesn't
- Update instructions after you spot a bad reply in Chat Logs - a one-line rule often fixes a whole class of mistakes
- Keep it short enough to scan - a long, sprawling instructions field is harder to keep consistent than a tight one
- Don't cram product facts in here - that's what training material and the Product summary are for; Instructions is about behavior, not knowledge
Product
What it is
The Product screen holds the Product name, the Language the agent works in, and an AI-generated summary of what your agent is actually talking about - the product, service, or webinar topic it's supporting. eChat builds the summary from the agent's training material.
What it's best for
- Giving the agent a clear, condensed sense of what it's representing, separate from the raw detail in its training sources
- Spotting when training material is vague or scattered - a bad summary usually means the source material needs work
- Keeping an agent focused if its training touches on adjacent topics it shouldn't be answering questions about
How it works
Once at least one training source is ready, click Generate Product Summary. Later, Re-generate Product Summary from Training Data produces a fresh summary from current training, or you can edit the text directly to correct or sharpen it by hand. This summary is part of what grounds the agent's understanding of the product, which is exactly why one agent should stick to one product. If you sell more than one, a single agent's Product summary and training blur together across them, which makes its answers less accurate for both. Create a separate agent per product instead, each with its own training and its own Product summary.
Best practices
- Regenerate the summary after a significant training update, like adding a new batch of webinars or a large document upload
- Read it after every regenerate - if it drifts or picks up stale details, edit it directly rather than letting it stand
- Treat a confusing Product summary as a signal to clean up training material, not just something to patch over by hand
Agent
What it is
The Agent screen sets the agent's identity - its name, avatar, and whether it's labeled as AI - along with which AI model it uses to generate replies, how strictly it checks its answers against training material, and how it works with your human moderators.
What it's best for
- Making the agent feel like a real member of your team in the webinar chat, not a generic bot
- Balancing reply quality against cost by choosing a cheaper, faster model or a higher-quality, more expensive one
- Tightening or loosening how carefully the agent double-checks itself before answering
- Deciding how the agent involves your moderators - flagging messages for them, leaving them notes, and whether it can hand a conversation to a person
How it works
Identity. Set a Name and an avatar (Upload image, or Replace image to change it). In an eWebinar webinar, this is exactly what attendees see in the live chat - the name and avatar become the moderator identity the agent posts under, alongside your human moderators. Turn on Show as AI to add an "(AI)" badge after the agent's name so attendees know they're talking to an AI; the label syncs to eWebinar. It only changes how the agent is labeled, not how it behaves.
AI Model. Choose the model used for visitor replies. The picker shows the credits each response costs: GPT-5.6 Luna (1), GPT-5.6 Terra (3, the default), Claude 4.6 Sonnet (3), Claude 4.6 Opus (5), Claude 4.8 Opus (6), and GPT-5.6 Sol (10). Higher-cost models trade more credits per reply for more careful, nuanced answers. See Choosing an AI Model.
Answer confidence. Set Grounding strictness to Creative, Balanced (the default), or Strict. This controls how aggressively the agent double-checks its own claims against training material before replying. Stricter settings mean more checking, so answers stay closer to what your training material actually says. At every setting the agent still answers in its own words - it never falls back to a canned "I don't know" - so tightening this doesn't risk the agent going silent, it just makes it more conservative about how far it stretches beyond what it was trained on.
Escalation. In eWebinar the agent is an AI moderator working alongside your human moderators, and this setting decides how far it involves them. Chatbot only: the agent works alone and never mentions your team. Leave a note: the agent can flag specific attendee messages as Needs reply for your moderators and leave them private notes, while it keeps helping the attendee itself; it never hands off. Human hand-off: everything Leave a note does, plus the agent can hand a conversation to a person when the attendee agrees or asks for one. With Human hand-off, the Human follow-up cards decide which situations can lead to a hand-off: Answer is uncertain, Frustrated language, and Human request (all on by default). Whether a hand-off is possible on a given webinar is controlled in eWebinar; if escalation is turned off for every linked webinar there, these cards are disabled and a notice explains why. See Working with Human Moderators.
Wait time (set in eWebinar). How long the AI waits for a human to answer first isn't an eChat setting - it's AI answers if a human doesn't answer in … sec in each webinar's AI moderator settings in eWebinar (0 to 900 seconds; Human hand-off agents start at 120, Chatbot only and Leave a note agents at 0). During the wait your moderators are notified and can reply first; if one replies or starts typing, the AI doesn't answer that message. The AI skips the wait when it's already the one talking with the attendee, or when no moderator is available. The wait only applies while attendees can request a person on that webinar. See the wait time for the details.
Best practices
- Start with GPT-5.6 Terra and Balanced grounding - they're the defaults for a reason and work well for most agents
- Move to Strict grounding for anything regulated or high-stakes, like pricing, refund policy, or medical or legal-adjacent claims, where a slightly-off answer is costly
- Use Creative sparingly - it helps when training material is thin and you'd rather the agent reason its way to a useful answer than stay narrowly literal
- Upgrade the model for a sales-facing agent where a sharper, more persuasive reply pays for itself, and keep support-style agents on a cheaper model where speed and volume matter more than nuance
- Pick a name and avatar that read as a real person or team identity - visitors respond better to "Ask Jamie" than "AI Assistant" - and use Show as AI if you want to be upfront that it's an assistant
- Only choose Human hand-off if moderators actually watch the chat; otherwise Leave a note gives your team a Needs reply list without promising live help
- If your team moderates live, keep a short wait in eWebinar so people answer first; for sessions nobody watches, set it to 0
Welcome Message and Visitor Settings
What it is
eChat also has Welcome Message and Visitor Settings screens (a greeting, and name and email collection). They belong to standalone chat agents, so they don't appear in the editor for agents linked to eWebinar.
How it works
For eWebinar-linked agents, eWebinar already knows who each attendee is and runs the chat room, so the agent doesn't send its own greeting or ask attendees for their name and email. If you're looking for these screens on an eWebinar agent, that's expected - there's nothing to configure.
Best practices
Put anything you'd have said in a greeting into Instructions instead (for example, what the agent can help with), and manage attendee details in eWebinar.