Atelier
A log of building infrastructure, systems taken apart, and the hammer blows that sharpen an engineer's thinking.
Now writing
The Native Era: Production AI Agent Architecture
Skip the hype. Six stops on how AI agent architecture actually changed in production — from prompts and RAG, through ReAct and multi-agent collapse, to harness engineering, zero-trust sandboxes, RL-native training, and the company as an operating system.
Everything, newest first
3 postsThe Native Era: Production AI Agent Architecture · part 3/6Who Stuffs the Text, Who Draws the PathFunction calling, LangGraph, Structured Outputs and a lesson from MoE: the model keeps the right to reason, but the right to turn a string of tokens into an action has moved to the harness, a layer of code with a graph, typed state, a runtime and an audit trail. The Native Era: Production AI Agent Architecture · part 2/6ReAct, the Cut String, and Outsized Expectations of a Technology Still Too YoungMRKL, ReAct, Toolformer, regex parsers, the apology loop, virtual departments and the unsigned tape, reread from the original papers, commits and issues: early agents saw the future clearly, but had to build it on a contract written in prose. The Native Era: Production AI Agent Architecture · part 1/6Ask Nicely, Stuff the Context… and Still Not Have a SystemGPT-3, InstructGPT, the original RAG, the geometry of embeddings and Lost in the Middle, read from the papers: asking nicely and stuffing the context change what the model sees, not what the system may do or how it proves it did it.
That’s everything in the Atelier so far.
© 2026 espresso.hoagg — all rights reserved