Enhancement isn’t a leap: it’s a ladder. You climb one step at a time — first what breaks, then what's inefficient, then output quality and only then the performance. The diagram shows this priority order, from the bottom (errors) to the top (speed).

Illustrative diagram — conceptual recreation of the enhancement order, not a real screenshot.

What it is: a workflow that “works” is basic and has no apparent errors; one that “works well” handles edge cases, provides a good user experience, and is efficient. The enhancement philosophy is to close that gap in layers — one change at a time, in steps, so you don’t get overwhelmed.

Key concepts

What it is: there's a clear order for enhancement— error → logic → output quality → performance. First, what breaks; then, what’s inefficient; then, the experience; and finally, speed. Skipping this order wastes time.

The four steps, in order

Key concepts

What it is: when the vector store can’t find it a policy for responding, the agent shouldn’t make anything up. The solution: the agent emits a confidence marker (“policy found” / “no policy”), a code node interprets it, a IF branches off, and a new polite draft says "let’s research this and get back to you".

agente → marcador: "no policy found"
   ↓
nó de código → lê o marcador
   ↓
IF → política encontrada?
   ├─ sim → rascunho normal
   └─ não → rascunho "vamos pesquisar
            e retornar" (bem formatado)

Key concepts

What it is: process emails only newer than X minutes, so you don't reprocess old inboxes. Instead of adding a time-check node, the AI discovers Gmail's search syntax and adds a filter “newer than” directly in the trigger.

# pedido em linguagem natural:
"Processe só e-mails dos últimos 10 minutos."

# a IA adiciona ao gatilho:
query: newer_than:10m
# sem nó separado de checagem de tempo

Key concepts

What it is: not every email deserves a reply. A code node after the trigger detects out-of-office messages, delivery notifications, newsletters, and no-reply/postmaster/bounce/marketing senders, add a boolean isAutomated and the reason, and branches to skip. This creates three gates: automatic? → support? → policy match?

# falha: header.toLowerCase() vs "No-Reply@..."
# corrige: comparar lower E original

Key concepts

What it is: adjust the tone of the responses to be friendly, helpful, casual, and enthusiastic. The change is made in two places: the AI agent's system prompt (technical answers) and the fallback draft, "let's look that up." Here we enter the output quality layer.

Key concepts

What it is: extract the customer's name from the signature/closing from the email (patterns like "Best, João" / "Thank you, Sara"), store it in a variable, and use it in the greeting—with a fallback “Hello” when there’s no subscription. A new code node handles the extraction; the agent prompt and the two drafts then use the name.

padrões: "Abraços, João" / "Obrigada, Sara"
   ↓
variável nome = "João"
   ↓
saudação: "Olá João," # ou "Olá" se vazio

Key concepts

What it is: a high/medium/low badge based on the score showing how well the knowledge base matched — for internal review, appended to the end of the draft. The agent emits the marker (e.g., high > 0.8; medium 0.6–0.8; low < 0.6); a code node interprets it and builds an HTML badge (green/yellow/red, gray when there's no match).