AI agents are systems problems
The difficult part of useful agentic AI is not the model call. It is everything required to make judgment and action reliable.
Insights
One publication. Many difficult questions.
Technical articles, research, engineering notes, and professional analysis about intelligent systems and the real engineering required to make them useful.
Topics describe what an idea is about. Formats describe what kind of piece it became.
Featured / opinion
AI agents are systems problems
Models matter. But architecture, context, permissions, evaluation, and operations decide whether an agent becomes infrastructure or remains a demo.
Read the insight2 insights
articleA practical guide to agentic engineering for DevOps practitioners, covering harness, context, prompts, RAG, memory, tools, control loops, sandboxes, guardrails, observability, and evaluation.The model may supply the reasoning, but the harness makes an agent operational. Here is how prompt, context, runtime, tools, memory, safety, and observability fit together—and why the stack should feel familiar to DevOps engineers.
articleA practical guide to tokenization and embeddings, covering BPE, WordPiece, Unigram, SentencePiece, word2vec, GloVe, fastText, contextual Transformers, and modern retrieval models.Tokenization breaks text into model-readable units. Embeddings place those units—or entire passages—into learned numerical spaces. Here is how the major approaches work and why the distinction matters.