AI adoption that stuck

I taught myself agentic engineering on the job at EasyPost, and it grew into a second role: helping the 40+ engineers around me get real value out of AI tools.

I started with context. I set up the seven or eight repositories our extended team worked in with detailed agents.md files, plus matching copilot-instructions.md files so GitHub’s automated reviews knew our conventions too.

Context across services was harder, because no single repository explains how the services fit together. That led to a RAG pipeline on Amazon Bedrock, where I did much of the technical design: small Haiku agents scan each repository and write structured markdown, which feeds a vector database that agents can search.

Tools only help if people use them well, so I ran weekly AI office hours and demos. What we learned in those sessions became the Agentic Toolkit, the team’s shared library of Claude skills, setup guides and example prompts.

The skills show the approach. A planning skill makes the agent research the code, tests and docs before proposing anything, and spell out correctness, data integrity and performance. A PR review skill treats Copilot’s suggestions skeptically and waits for approval before changing code. A Sentry skill treats error data as untrusted input.

An anonymized copy of the toolkit is public at github.com/lukemiller/agentic-toolkit.