Page 2 of AI Engineering
A technical teardown of the agent loop, context assembly, and tool wiring inside Claude Code, Cursor, and Codex — the machinery that makes modern coding agents work.
GLM-5.3 Flash is more than a cheap coding model; its native multimodal training makes it practical for video, 3D, CAD, slides, financial earnings, Godot games, vision-driven UI, and computer-use work. This guide walks through the recommended prompt patterns the Z.ai docs use for each non-coding…
A troubleshooting-oriented look at five anti-patterns that quietly degrade LLM output — from stacking temperature and Top-P to misusing repetition penalties on reasoning models — with the reasoning and a safer alternative for each.
A practical walkthrough of Argilla's FeedbackDataset data model — Fields, Questions, Guidelines, and Records — aimed at engineers assembling preference data for DPO and RLHF.
A practical engineer's explainer on GPT-6 Astra's recurrent depth architecture, the looped-transformer design it implies, and why the chain-of-thought the API returns is now a compressed summary rather than a transcript.
A developer's guide to running open-source LLMs on Apple Silicon, explaining how Ollama, llama.cpp, and MLX-LM work together and why unified memory changes the math for local inference.
An architecture-focused comparison of supervisor, peer-to-peer, adversarial, and graph-based multi-agent coordination patterns, with trade-offs, failure modes, and selection criteria.
GLM-5.3-Flash activates only 18B parameters per token, but all 320B weights must stay resident in VRAM — here is why 'active' and 'resident' are different things.
Headline math, three-week test results, and the burst-workload edge cases that decide whether switching to the GLM Coding Plan actually saves money versus Claude Max or Cursor Pro+.
A deep dive into the Memory MCP reference server's knowledge-graph persistence model, comparing it to vector stores and showing how to wire it into a client for long-term agent context.
A step-by-step Sol→Astra migration playbook covering eval set construction, acceptance-gate scorecards, and prompt rewrites tuned to Astra's behavioral tendencies.
Caveman, Headroom, and Graphify attack different parts of the token-cost problem — output voice, incoming context, and repeated file reads — so they stack cleanly when configured with attention to layer order and KV-cache prefix stability.
A deployment-and-scale-driven comparison of Qdrant, Pinecone, Weaviate, Milvus, and Redis for production RAG, with a decision framework for teams choosing at the selection stage.
A blueprint for orchestrating multiple LLM agents — Diversity, Privacy, QA Curation, expert deliberation, reasoning models, and label-targeted generation — to build custom synthetic QA evaluation sets for RAG systems.