Dr. Strangeagent, or:
How I learned to stop prompting and love the state machine
AWS Meetup, Dublin; 24.03.2026
Thank you for joining us
ABSTRACT
AI agents nail demos and drift in production. They skip steps, hallucinate results, and present confident garbage. The fix is not better prompts – it is better architecture.
This talk shows how Step Functions patterns – typed I/O, explicit transitions, choice states, parallel execution, error catchers – solve the reliability problem for AI agents. I will walk through a real mutation testing workflow (Chaos Monkey for your test suite) that orchestrates five parallel AI agents, routes on data, and validates intent at the terminal state, without the orchestrating agent remembering anything.
Three ideas that land:
-
Split intelligence from execution. AI reasons freely to scope the problem, then a typed state machine drives the steps. Same mental model as Step Functions, applied to agents.
-
Isolated sub-tasks. When a step needs real thinking, spin up a fresh agent with focused context and only the data it needs. The main thread stays deterministic. The sub-task gets the model's full attention.
-
Self-driving I/O. Every step's output carries the next instruction. The agent follows the chain - it does not maintain a plan.
If you have used Step Functions on AWS, you already understand the mental model. This talk shows where it maps directly to AI orchestration and where the AI-specific patterns diverge.
Takeaways
-
Why agents drift and why prompt engineering cannot fix it
-
A concrete architecture you can implement the following week
-
Some core implementation patterns that have been tested in anger
BIOGRAPHIES
Glen Keane
Glen is a senior software engineer at Brightbeam who wears many hats. Lately, he’s been focused on building intelligent systems using AI. In his past he built smart systems without it.
A fullstack developer in all the worst possible ways - he does backend, frontend and devops in both startup and enterprise environments. Previously, he’s been a node.js core maintainer at Nearform; the first engineering hire (outside the founding team) in an Irish Unicorn, Wayflyer; and a Head of Engineering for a local startup.
He likes to pretend he’s cosplaying as Chewbacca when he needs a haircut (which is most of the time) and enjoys sitting by beaches in the sunny south east writing code and sipping coffee.
Brian Finnegan
Brian brings three decades of experience to his role as a Principal Solutions Director at Brightbeam. His expertise in technology delivery stems from a diverse background – which includes previous roles as a Technical Account Manager at AWS and Solutions Architect at Sungard and Presidio. As well as Cloud Architect at Zurich Insurance.
This broad perspective gives Brian a pragmatic edge (cutting and otherwise) in bridging the gap between the promise of AI and other new technologies and the gritty reality of scaling it in production.
Blake Rizk
Blake is Head of AI and Transformation at Laya Healthcare, part of the AXA Group, where he leads the strategy and delivery of AI-driven transformation across the organisation. With over 20 years of experience in technology, data and digital transformation, he focuses on applying practical AI solutions to modernise core healthcare insurance operations and deliver measurable business value.
Blake is currently responsible for advancing the adoption of artificial intelligence across many areas including claims processing, document intake, customer service, and knowledge management.
ABOUT BRIGHTBEAM
Brightbeam integrates digital intelligence into complex and regulated enterprise environments. What we like to call ‘the hard stuff’. Our mission is to the most helpful services company in the world and sitprovide end-to-end support through four core service pillars:
Focus: strategy & prioritisation
Defining roadmaps and prioritising high-impact opportunities. This includes Very Rapid Prototyping of applications to de-risk innovation, validate concepts and prove strategy can be turned into working solutions
Build: implementation & integration
Developing and deploying production-ready, compliant AI that integrates seamlessly into existing technology stacks and operational workflows.
Embed: coaching & change management
Transforming organisations via workplace education, culture and change management to the rapid adoption of AI-native ops models. We create the knowledge and focal points to embed AI at previously impossible rates.
Invent: research & development
We partner with forward-thinking organisations – clients and other vendors – to explore the cutting edge of AI applications.
WE'RE RECRUITING
CONNECT WITH US
LinkedIn networking
Join the conversation on LinkedIn and connect with us. Share your takeaways, ask questions, and continue collaborating on Second Nature Manufacturing.
