ISOM 260 · Session 8
Multi-Agent Workshop
45:00
Build the Researcher Agent
15 min · 45:00 – 30:01
- Define the Researcher's system prompt
- Set structured output format (Key Facts, Statistics, Trends, Confidence)
- Test on "AI agent enterprise adoption trends"
Build the Analyst Agent
15 min · 30:00 – 15:01
- Connect Researcher output as Analyst input
- Define pattern identification prompt
- Generate insights + recommendations
Add Editor + Run Full Pipeline
15 min · 15:00 – 0:00
- Build Editor agent (fact-check + clarity + score)
- Wire all three agents into run_pipeline()
- Run end-to-end and compare with single agent
Open Notebook in Colab
Phase 1 — Researcher
- Start with a clear role:
"You are a specialized Research Agent" - Force structured output with
"Return EXACTLY this format:" - Include a confidence level for accountability
Phase 2 — Analyst
- Feed
researcher_outputdirectly into the analyst's user message - Tell the analyst:
"Don't just restate — find what the researcher missed" - Require specific fact IDs to prevent hallucination
Phase 3 — Editor + Pipeline
- Editor receives BOTH researcher and analyst output
- Use a quality score 1–10 for measurable feedback
- Compare: Is the pipeline output really better than single agent?