JS Nation 2026 Agent Runtime Workshop

This folder contains runnable code samples that map to the workshop plan. The practical flow is built around an email triage agent that classifies incoming emails into task, event, or no_action and routes them to downstream agents.

Step 01 uses a direct OpenAI-compatible HTTP call with no SDK, and falls back to naive keyword routing when OPENAI_API_KEY is not set. Step 02 uses the Google ADK (@google/adk) for one classifier agent. Later steps build specialist agents and orchestration around that classifier.

Quick start

cd 2026-02-28-jsnation-workshop
node -v
npm install
npm run start:01

Mapping to the agenda

1. Agent theory. What they are in general
   • See src/runtime/README.md for the runtime model and architecture notes.
2. Developing a simple agent
   • Run npm run start:01 and open src/01-standalone/run.js.
3. Agent SDK theory
   • See src/02-sdk/README.md for the Google ADK overview.
4. Developing an agent using the SDK
   • Run npm run start:02 and open src/02-sdk/run.js.
5. Orchestration theory. What to do when there are multiple agents
   • Run npm run start:03 and open src/03-orchestrator/run.js.
6. Developing one or two more agents and integrating them into n8n
   • Run npm run start:04 and open src/04-n8n/README.md.
7. Theory of general orchestration rules. Security, monitoring, etc.
   • Run npm run start:05 and open src/05-security-observability/run.js.
8. Practice
   • Use src/examples/ for exercises and tasks.
9. End

Structure

• src/01-standalone/ raw LLM email triage over HTTP without an SDK
• src/02-sdk/ Google ADK classifier agent
• src/03-orchestrator/ specialist agents plus monolithic orchestration over the step 02 classifier
• src/04-n8n/ n8n integration nodes for visual orchestration over the step 02 classifier and step 03 specialists
• src/05-security-observability/ wrappers for prompt-injection checks and success/error observability
• src/examples/ practice tasks and sample emails
• src/runtime/ runtime model notes (kept for theory section)
  • Also includes the small OpenAI-compatible helper reused by step 01

Notes

• Step 01 uses a real model when OPENAI_API_KEY is set, otherwise it falls back to naive keyword routing.
• Step 02 stays on the ADK code path even without credentials by using a local keyword-based BaseLlm fallback.
• Step 02 also supports Gemini, OpenAI, or a local OpenAI-compatible endpoint via SDK_PROVIDER; see src/02-sdk/README.md.
• Step 03 is intentionally monolithic and adds specialist agents around the step 02 classifier.
• Step 04 reuses the step 02 classifier and the step 03 specialists in n8n nodes, plus the step 03 router as a shortcut node.
• Step 05 wraps the step 02 classifier and step 03 specialists for prompt-injection blocking and success/error monitoring instead of re-implementing routing.