5 AI Agent Reliability Patterns, Taught by a Cyberpunk Cat RPG

I built a cyberpunk cat RPG for AWS Summit LA where you play a TabbyWarrior, SiameseMage, MaineCoonPaladin, or SphinxRogue fighting through Neo-Pawsburg in 2087. Next to the game, you can watch the real production AI architecture running underneath it live: Bedrock, Step Functions, Lambda, DynamoDB, the whole thing. This episode is the architecture tour, minus the glazed-eyes-at-a-diagram problem. I walk through the five reliability patterns that decide whether an AI agent survives production instead of just surviving a demo: grounding the model with retrieval before it generates anything, retries and idempotency that don't break randomness, the boundary where the model proposes an action and the system decides whether to execute it, structured observability built in from the first Lambda invocation, and state that lives in a database instead of inside a prompt. Then I tell you about the day I almost paid seven hundred dollars a month to keep a search index warm for a game about cats, and what AWS announced the very next day. 00:00 Cold open: Neo-Pawsburg, 2087 02:33 Why cats, why cyberpunk, why AWS Summit LA 5:33 The five boss fights: grounding, retries, who's in charge, observability, state 11:01 The seven hundred dollar serverless lie 14:20 What I left out on purpose (and why the patterns still hold) 17:18 Closing If you build agentic AI systems and want the actual reliability patterns without the vendor slide deck, or if you've ever wanted someone to explain what "the AI just makes an API call" is quietly doing on your behalf, this episode is for you. An AI agent is a distributed system. It plays by the same rules. The cats just make the rules easier to sit through. Repo: https://github.com/chaotictoejam/neon... Chaotic Commits, Episode 13 feat: I built a cyberpunk cat RPG to teach production AI architecture #AIarchitecture #AWSBedrock #distributedsystems #softwareengineering #AIagents #serverless #observability #productionAI