How Much of Your AI Agent Should Be Code?
Harness engineering in practice: how much of an AI agent to hand to code depends on what a wrong answer costs. Three of our own agents, from one end to the other.
Dr. Phil Winder
/
8 minArticles in this topic
Dr. Phil Winder
/
8 min
Copilot Studio vs Azure AI Foundry vs Build: What One Document Costs
Copilot Studio vs Azure AI Foundry vs building your own, costed on one real 236-page report: $725 a document against $37, and where the line falls.
Dr. Phil Winder
/
17 min
vLLM vs Ollama vs SGLang: LLM Inference Compared 2026
vLLM vs Ollama vs SGLang on one H100 with Qwen3.8-27B and Llama 3.1 8B: throughput, latency and cost per million tokens, harness included.
Dr. Phil Winder
/
11 min
DeepSeek Harness vs OpenCode: Which One We Run in 2026
DeepSeek Harness vs OpenCode 2.0, compared from running both in our own fleet: architecture, model support, cost, containment and which one suits a team.
Dr. Phil Winder
/
13 min
A Comparison of AI Agent Harnesses in 2026
A comparison of AI agent harnesses in 2026: Claude Code, Codex, OpenCode, Qwen Code, DeepSeek Harness, Goose and Zed Agent, plus when you need a framework.
Dr. Phil Winder
/
17 min
How to Build an AI Agent in 2026: Frameworks and Working Code
How to build an AI agent in 2026: Pydantic AI vs LangGraph vs build-your-own compared, two worked examples with real code, and the evaluation and memory patterns that survive production.
Dr. Phil Winder
/
11 min
RAG vs Fine-Tuning in 2026: A Decision Framework for LLM Teams
An opinionated 2026 decision framework for choosing between retrieval-augmented generation (RAG), fine-tuning, and hybrid approaches for LLM applications. Decision tree, comparison table, and named tooling.
Dr. Phil Winder
/
9 minNewsletter
AI engineering insights, in your inbox
Case studies and industry analysis from our team. No hype, roughly monthly.