SRAL: A Framework for Evaluating Agentic AI Architectures
State, Reason, Act, Learn: a four-part lens for evaluating whether an agent's architecture survives production, not just the demo.
Read the frameworkFounder, Foydy Former VP of Services and Solution Architecture, Akka (formerly Lightbend) 14 years of industry experience creator of the SRAL framework
I design and evaluate agentic systems where architecture, not the model, decides what survives production.
Currently building Foydy, a consumer-insights company in India: we tell brands whether a product will sell, from real households using it at home for two weeks. I also write about AI systems that think, act and recover under pressure, and advise a few teams on getting them into production.
10 years writing105 essays1 frameworkDOI'dopen-source
State, Reason, Act, Learn: a four-part lens for evaluating whether an agent's architecture survives production, not just the demo.
Read the frameworkThe framework I evaluate every agent with: State, Reason, Act, Learn. The mental model the rest of the blog builds on.
How AI shopping agents break the measurement commerce runs on.
The distributed-systems idea under most of my writing: a system's report of itself is a claim, not a measurement.
Why your AI agent keeps repeating the same mistakes. Learning is distilling what to do differently, not storing what happened.
My core thesis in full: the runtime, not the model, is the agent.
Agent cost read as a scheduling problem, admission control and backpressure, not a billing dashboard.
What "consumer behavior" even measures once an agent, not a person, makes the purchase.
The vector-database failure modes that only surface in production, and what to design for.