Blog

Engineering notes on applied AI

No hype, no hand-waving. Practical writing on evaluation, retrieval, guardrails and the engineering that turns models into dependable systems.

Featured · Research

Giving AI agents memory that actually persists

Vector stores give agents recall, but durable memory needs summarization, decay and clear write policies too — otherwise context windows fill with noise instead of signal.

DStarix TechnoJune 30, 20268 min

Read article

Latest writing

All articles

Research

The ReAct pattern: how reasoning and acting changed LLM agents

The 2022 ReAct paper (Yao et al.) showed that interleaving reasoning traces with tool calls made language models dramatically better at multi-step tasks. It still underpins how most agents are built today.

DStarix TechnoJun 18, 20269 min

Engineering

Model Context Protocol: a shared standard for connecting agents to tools

Anthropic's Model Context Protocol gives agents a common way to discover and call external tools, files and data sources — replacing one-off integrations with a single open protocol.

DStarix TechnoJun 5, 20268 min

Engineering

Why most AI agent demos fail in production

A slick agent demo and a dependable production agent are different engineering problems. Evaluation harnesses, guardrails and observability are what actually close that gap.

DStarix TechnoMay 22, 202610 min

Fundamentals

What are AI agents, really? A practical definition beyond the hype

An agent isn't just a chatbot with a system prompt. It's a model wired to tools, memory and a control loop that perceives, plans, acts and observes — here's what that actually means in practice.

DStarix TechnoMay 8, 20267 min

Engineering

Multi-agent orchestration: LangGraph, AutoGen and CrewAI compared

When a single agent can't hold an entire task, orchestration frameworks let specialized agents plan, delegate and hand off work to one another. Here's how the major frameworks actually differ.

DStarix TechnoApr 24, 202611 min

The DStarix dispatch

One sharp idea on production AI, every other week.

No fluff. Just the evaluation, architecture and cost lessons we learn shipping systems — straight to your inbox. Unsubscribe anytime.

Have a problem worth writing about?

Bring us your hardest AI use case. We'll map a path to production — and maybe write the field notes together.

Start a project See our work