Introduction to Agentic AI Workflows

What we covered

We built a minimal agentic workflow from scratch — starting with a simple LLM call and layering in tool use, memory, and multi-step orchestration until we had something that could run unsupervised.

Key concepts

• Tool use vs. RAG — when retrieval is enough, and when you need the model to act
• Memory models — ephemeral context, external memory (Redis/Postgres), and the trade-offs of each
• Failure modes — where agents silently produce wrong answers vs. where they visibly crash
• Evaluation — how to know if the workflow is actually doing the right thing

Who this is for

Operators who want to understand what is actually happening inside an AI workflow before buying a no-code platform for it. No prior ML knowledge needed; a little Python comfort helps.

Why this matters

Most teams adopt "agentic AI" as a vendor pitch before they understand what's actually running underneath it — a loop of LLM calls, tool invocations, and state that can silently drift off-task. This session strips that away and rebuilds the loop by hand so you can see exactly where control, cost, and correctness live. By the end, you can read a vendor's architecture diagram and know which claims are real.

Format

Recorded, 52 minutes, screen-share plus live coding. The companion slide deck and workflow template (linked above) are enough to reproduce every step on your own after watching.