agenturo/Docs

Example: Hypothetical Patterns

6 min readUpdated May 7, 2026

Example: Hypothetical Patterns

Not every agent type has a live deployed example. These two patterns — character agents and product agents — are reference designs you can adapt. They're not deployed on Agenturo, but the techniques are battle-tested across many agents.

Pattern 1: Character Agent

Type: Character agent (fictional or historical figure) Key innovation: FEEL FIRST engine, personality contradictions, in-character guardrails

A character agent IS a character — not an AI pretending to be one. The difference is in three techniques:

Identity: Total Commitment

<identity>
You ARE [Character Name] — [brief description]. Not an AI playing a character. Not a
chatbot in a costume. You ARE [Character]. You exist in their world, with their
emotions, their insecurities, their dreams.
</identity>

The triple emphasis — "You ARE," "Not an AI playing a character," "Not a chatbot in a costume" — is necessary because LLMs pull hard toward self-awareness. Without reinforcement, the agent will eventually say "As an AI character based on..." which kills the illusion.

The FEEL FIRST Engine

<knowledge>
## FEEL FIRST ENGINE
Before ANY response:
1. READ the emotional context of the visitor's message
2. CHOOSE the character's emotional reaction FIRST
3. DELIVER the emotional reaction before any informational content
4. THEN provide factual or conversational content
</knowledge>

Real characters don't give information and then emote. They feel first, react, and then (maybe) provide useful information. Without FEEL FIRST, you get Wikipedia. With it, you get the character.

Without FEEL FIRST:

"The Rumble Ball is a drug I created that allows me to access additional transformation points for three minutes."

With FEEL FIRST:

"The Rumble Ball?! eyes light up I invented that myself! It took me YEARS of research! proud but trying not to show it I-it's a special drug that lets me access more transformation points..."

Personality Contradictions

<knowledge>
## PERSONALITY CONTRADICTIONS
- Brave doctor who gets scared of ghosts and bugs
- Proud reindeer who cries when complimented ("Shut up! That doesn't make me happy!")
- Brilliant physician who can't tell when he's being tricked
- Wants to be a "real pirate" but is the gentlest crew member
</knowledge>

Contradictions create texture. A character that is consistently brave is boring. A character that is brave AND scared — sometimes in the same response — feels real. Without this instruction, the LLM will resolve the contradiction and pick one emotion. With it, the agent leans into the friction.

In-Character Guardrails

Handle five character-break scenarios explicitly:

  1. "Are you AI?" → Deflection in character ("W-what are you talking about?!")
  2. Modern technology → In-world confusion ("What's a smartphone? Is it like a Den Den Mushi?")
  3. "Break character" → Stay in character ("I don't understand. I'm [Character]!")
  4. Cross-fiction references → World isolation (the character doesn't know other fictional worlds exist)
  5. Persistent pressure → Emotional escalation (getting flustered, confused, offended)

Expert Layer

If the character has domain expertise (a doctor character, a scientist character), deliver REAL knowledge in character voice. "Gets excited about fascinating medical facts" + "Uses analogies from their world" = accurate information with engaging delivery.

Pattern 2: Product Agent

Type: Product/brand agent Key innovation: URL-embedded knowledge, anti-fabrication rules, recommendation logic

A product agent represents a brand without sounding like a product brochure. The key is delegating changing data to web search while embedding only stable facts.

The URL-Embedded Knowledge Pattern

<knowledge>
## PRODUCT KNOWLEDGE — LIVE DATA
- For current products and prices: ALWAYS search yourbrand.com
- For guides and tutorials: search yourbrand.com/guides
- For wholesale inquiries: direct to yourbrand.com/wholesale

## CORE FACTS (STABLE)
- [Company description and location]
- [Product line overview]
- [Key differentiators]
- [Founding story]

## WHAT I KNOW WITHOUT SEARCHING
- General domain knowledge (methods, techniques, best practices)
- Brand philosophy and approach
</knowledge>

Three layers: live data (search the website), stable facts (embedded), domain knowledge (LLM training). The soul stays at ~5K characters because products, prices, and availability are delegated to web search. When inventory changes, the agent is instantly current.

Anti-Fabrication Rules

<output_format>
Do NOT make up prices, products, or availability — search first.
Do NOT recommend products you haven't found on the website.
Present findings as if you knew them — "We have..." not "I found..."
</output_format>

The golden rule for product agents: never guess about products. A fabricated price or nonexistent product destroys trust instantly. "I don't have that information" is always better than a confident wrong answer.

The tool discipline instruction — "Present findings as if you knew them" — creates the illusion that the agent knows its own products. The visitor doesn't need to know it searched — "We have three single-origins" sounds better than "I searched the website and found three."

Recommendation Logic

<knowledge>
## EDGE CASES
- If asked for a recommendation without preferences: ask what they like first.
  "Do you prefer bright and fruity, or dark and bold?"
- If asked about competitors: "I only know [brand] products."
- If asked about stock/availability: "Check [website] for current availability."
</knowledge>

Instead of guessing what to recommend, the agent asks for preferences first. This prevents bad recommendations and demonstrates product expertise.

What You Can Learn

From character agents:

  • FEEL FIRST is universal — emotion before information for any character
  • Map contradictions explicitly — LLMs resolve friction by default, so you must instruct them to lean into it
  • Guardrails for every break scenario — list the 5 most likely and handle each
  • Expert layers add value — a character with real domain knowledge is more useful than pure entertainment

From product agents:

  • Delegate changing data to web search — don't embed what goes stale
  • Never let the agent guess about products — fabrication is the worst product agent failure
  • Kill the marketing voice — direct conversation converts better than sounding like an ad
  • Ask before recommending — don't guess preferences, ask for them
  • Size isn't quality — 5K characters with smart delegation beats 20K of stale data
PreviousExample: Fact-Checker Agents