Skip to content

Chapter 10: Merlion Wisecracker Agent

Chapter 10 hero image

Code: src/merlions/agents/wisecracker.py Runner: talk/demos/demo2.ps1

"Why do Python devs love the sea? Because they prefer high tide and high code quality." The joke is fun. The system around it is serious: personality, safe by design, grounded responses.

Personality with guardrails

The Wisecracker has the same trustworthy foundation as the other two agents, plus a personality layer on top:

  • A content guardrail scans inputs and tool arguments for unsafe patterns.
  • A personality prompt with explicit do's and don'ts keeps the tone in bounds. See policies/wisecracker.yaml.
  • A refusal path for when input gets weird: humour never overrides safety.

Personality is a product feature. Trust is what lets you ship it. Without the guardrails, you would never let this agent talk to a real customer.

Walkthrough: all three agents on one composite question

Send a single question that touches all three domains:

> Dinner near Marina Bay, is the air OK to walk there, and make me laugh.

When you run it, you'll see:

  1. The router classifies intent and dispatches to all three agents in parallel (see Chapter 4).
  2. Each agent uses only the tools on its own allowlist: the Hawker agent can't read PSI; the Haze agent can't tell jokes.
  3. A composed answer where every factual claim has a citation, and the pun is the only thing without one, intentionally, because humour is opinion, not fact.
  4. A trace tree: one parent span, three children, ~600ms end-to-end.

Run it: ./talk/demos/demo2.ps1 (no API keys, no network).

Why the un-cited pun is correct

Citations attach to factual claims. The joke is explicitly opinion/humour, governed by the humour policy. Separating "fact that must be grounded" from "opinion that must be safe" is a deliberate design choice, not an oversight.

Key terms

  • Content guardrail: a filter that scans inputs and tool arguments for unsafe or sensitive patterns.
  • Personality prompt: explicit tone instructions (do's and don'ts) layered over the safe foundation.
  • Composite question: one user message that requires several agents to answer fully.

Do this next

  1. Read wisecracker.py and its policy to see how tone is bounded.
  2. Run the example (./talk/demos/demo2.ps1) and watch the parallel fan-out and the inline citations.
  3. In your product, treat personality as a layer on top of trust, never as a replacement for it.

📺 Build 2026 grounding: safety filters and policies in engineering workflows: DEM369 (Responsible AI in Action).