Agentic Workflows: Stories from the Engine Room
summary
Has built several dozen agentic workflows for banking and insurance and finds most succeed only at level 1–2 (in‑context LLMs with short memory); recommends starting with one agent per process rather than a “digital employee,” limiting agents to no more than six tools, and using the CO‑STAR prompt structure plus 10–50 examples to stabilize outputs. Prototypes use simple infra (Google Sheets, JSON validation loops), always capture every prompt for tracing, require human approval on high‑risk steps, and monitor API costs and integrations (Zendesk example) to keep adoption and ROI.
AI-written summary · about
chapters (27)
- 0:00Introduction and background
- 1:27Why agentic workflows matter
- 2:51Levels of agent complexity
- 4:57Stories from real projects
- 5:17Pitfall one: Digital employee illusion
- 6:39Pitfall two: Automating wrong processes
- 7:39Pitfall three: Over-engineering agents
- 8:00Pitfall four: Prompt chain trap
- 9:01How prompt chaining works
- 12:11Example of agent collaboration
- 12:32Pitfall five: Overloaded agents
- 13:16Pitfall six: Prompting difficulties
- 15:02Structuring prompts effectively
- 18:06Prompt design best practices
- 19:52Temperature and model parameters
- 21:18Blackbox problem and explainability
- 22:20Human in the loop necessity
- 23:44Grounding results and reducing hallucinations
- 24:24Reliability is your responsibility
- 25:26Debugging agent loops with NA
- 26:29Forgotten integration pitfalls
- 27:31AI adoption and resistance
- 28:56Unexpectedly high costs
- 29:37Avoiding obsolete architectures
- 31:41The bitter lesson in AI
- 32:22Recap and key takeaways
- 33:26Q&A and closing remarks
related talks
50mAI Agentic Workflows & Design Patterns
26mFour Steps to Embed AI Agents into Workflow
26mFour Steps to Embed AI Agents into Workflow
16mAgents vs Workflows: Why Not Both?
50mUnpacking an Agentic Workflow
26mBuilding Agentic Workflows: Patterns for Orchestrating Intelligent Systems.
14mOrchestrating Agentic Workflows with Governance and Trust
57mWhen AI Agent Teams Work (And When They Just Create More Work)