Configuring an agent
Persona, model settings, versions, and publishing safely.
Persona
The system prompt is the single most consequential field in Vicero. It is read before every conversation and it decides what the agent is.
A prompt that works tends to say four things:
- Who the agent is and who it serves — "You answer questions from customers of Acme about their orders."
- What it must not do — topics to refuse, promises it cannot make, the fact that it must not invent policy.
- What to do when it does not know — name the escape hatch: say so, and offer a human.
- Tone — in a sentence, not a paragraph.
Alongside it: display name and avatar (what visitors see), a welcome message, suggested prompts to give people a way in, and a fallback message.
Model settings
| Setting | What it does | Suggested |
|---|---|---|
| Provider / model | Which LLM answers | Groq to start — fast and free-tier |
| Temperature | Randomness | 0.2–0.4 for support; higher only for creative work |
| Max tokens | Cap on reply length | Enough for a complete answer, not an essay |
| Top-p | Nucleus sampling | Leave alone unless you know why |
| Stop sequences | Force an early stop | Rarely needed |
Temperature is the one people get wrong. A support agent at 0.9 will phrase the same
policy three different ways to three customers, and one of those phrasings will be wrong.
Provider keys live under Settings → Provider keys, and can be set per organization or per agent. An agent pointed at a provider with no key configured fails its turns outright.
Toggles
Per agent, you control whether it can:
- search the knowledge base (RAG),
- call tools,
- keep memory across a conversation,
- hand off to a human.
Turning off what an agent does not need is worth doing. Each capability is an additional way for a turn to go wrong, and an agent that cannot call tools cannot be talked into calling one.
Versions
Every agent has a draft you edit, and — once published — a live version visitors talk to. They are independent: editing the draft never changes what customers see.
- Playground always runs the draft.
- Widgets and channels always run the published version.
- Publish promotes the current draft.
- Rollback re-publishes an earlier version.
This is why publishing is a separate permission. An editor can change everything about an
agent and still not affect a single customer until an admin or owner agrees.
Testing before you publish
Use the playground, and be adversarial about it:
- Ask something the knowledge base does not cover. Does it decline, or invent?
- Ask for something it should refuse. Does the refusal hold when you push twice?
- Ask an ordinary question. Is the answer cited?
- Ask in the sloppy way real customers write — no punctuation, half a sentence.
You can also save these as tests on the agent and re-run them before each publish, so a prompt change that quietly breaks an old answer shows up before your customers find it.
Templates and duplication
Templates give you a starting configuration for a common role. Duplicate copies an existing agent — the fastest way to try a significantly different prompt without risking the one that works.