How to Train 50 Engineers on Agentic AI in 4 Weeks (Without Slowing Delivery)

We've run this program three times now across three different engineering orgs — a total of 140 engineers. Here's what works and what doesn't.

The mistake: teaching tools instead of patterns

The first version of our curriculum was tool-heavy: here's how LangChain works, here's how to use the Anthropic API, here's how to build a RAG pipeline. Engineers got to the end knowing a lot of APIs and none of the judgment calls.

The second version organized everything around patterns: tool design, eval-driven development, escalation architecture, RAG pipeline design. The tools became examples. Engineers got to the end knowing when to use each pattern and why.

Week 1–2 structure

Day 1: What makes an agent different from a chain. Tool design principles. The escalation pattern. Hands-on: build your first production-quality tool.

Day 2: RAG pipeline architecture. Chunking, embedding, retrieval. Eval-driven development: write your test cases before your agent. Hands-on: build a RAG-backed Q&A agent over a real document corpus.

Weeks 3–6: The project

Every engineer ships a real agent — not a tutorial project. We ask them to pick something from their actual work: a support agent, an internal tool, a data query agent. They have 4 weeks to get it to production quality (which means: it has an eval harness, an escalation path, and observability).

What accelerates the cohort

The single biggest factor: engineers helping each other. We create a shared Slack channel and a weekly group session where people show what they're building. The peer review dynamic is more valuable than any lecture.