From Agent Demo to Agent System: A Production Architecture Workshop

​Most AI agents look great in demos and break in production. This workshop walks through the architecture that keeps them alive — not theory, working code, real failure modes, and a reference repo you can clone.

​What you get

• ​Hands-on walkthrough of production-ready AI agent architecture

• ​Nine harness components taken from a broken baseline to a fully wired system

• ​Architecture you can adapt to your own stack

​Who this is for

• ​Engineering leaders shipping AI agents, or planning to

• ​Directors, VPs, senior managers, staff engineers making build-vs-buy calls

• ​Teams already running agents in dev and hitting production blockers

• ​Anyone whose finance team asked "what does one AI ticket cost?" and could not answer

​Not for

• ​Researchers chasing SOTA benchmarks

• ​ML scientists tuning models

• ​Folks looking for prompt engineering tricks

​What you will learn

• ​Why agents break in production — seven concrete failure modes and what each one costs

• ​How to build a harness around any LLM — nine components: context engineering, tool access, orchestration, guardrails, memory, cost controls, sandboxing, observability, evals

• ​Why the harness is the moat — model quality is commoditizing; enterprises are deciding build/buy and Harrison Chase built Deep Agents reverse-engineering Claude Code patterns

• ​Two-layer permission model — guidance in the prompt, enforcement in infrastructure. Prompts get jailbroken; servers don't

• ​Deterministic vs probabilistic safety — why compliance teams won't accept "it works 80% of the time"

• ​Memory architectures — vector vs graph vs hybrid, when to use each, and where the industry is heading

• ​Cost controls — model routing that cuts 60–70% of spend, budget caps that stop runaway loops, circuit breakers for failing providers

• ​Observability patterns — Langfuse traces, LLM-as-judge evals, how to debug

• ​Migration playbook — how to add harness components to agents already in production. Order matters: observability first, then safety, then quality

​What you will leave with

• ​Architecture diagrams for all nine components

• ​A decision framework for framework selection (Claude Agent SDK, OpenAI Agents SDK, LangGraph, Deep Agents, CrewAI)

• ​The full ecommerceSupportAgent repo to clone and run

• ​A ramp-up guide covering every concept for deep reading

​Format

• ​15 min — business case and broken agent demo

• ​45 min — walk through the four highest-impact components: context engineering, guardrails, cost controls, observability

• ​15 min — full demo of the complete system

• ​15 min — Q&A and what your team should do next

​Live on Zoom. Recorded. Replay sent after. Free.