Chapter 13: Where to Go From Here

Three takeaways to carry home.

Three takeaways

1. Specialise your agents. Small, focused, governed. Not one mega-agent, but several little ones that collaborate. Smaller agent, smaller blast radius.
2. Build trust in, not on. Transparency, safety, reliability, observability, designed in from line one. Retrofitting trust is ten times more expensive.
3. Ship to Azure with observability from day one. Container Apps, Cosmos, Key Vault, Application Insights. A boring stack with reliable outcomes.

If you are a developer

Do this in your next week. It changes how you build forever:

• Clone the repo and run the three examples (talk/demos/README.md).
• Ship one tool with a typed signature and input validation.
• Add one guardrail as policy-as-configuration (see policies/).
• Write one eval from a real failure (see evals/).

If you are a decision maker

Go back to your team and ask three questions. If you don't get clear answers, you've found your roadmap:

1. Where are our agents' guardrails?
2. Where are our traces?
3. Where are our evals?

The closing thought

Singapore works because everyone respects the system: the queue at the hawker stall, the rules of the MRT, the trust that lets a million strangers cooperate every day. Our agents need the same thing: personality, yes; speed, absolutely; but trust is what makes them useful at scale.

Go build agents that are friendly, fast, and trustworthy. Singapore style. Lah, let's build. 🦁

Where to go next

• This repo: README.md, the three agents in src/merlions/agents/, and the eval suite.
• The chapters: revisit Chapter 6 (Trust), Chapter 11 (Azure), and Chapter 12 (Observe) as your build checklist.
• Build 2026 sessions: the full list with codes is in talk/context.md.

📺 Build 2026 grounding: DEM303 (Late to agentic coding? Don't panic, build.), the encouragement to just start.