Guides
Working with AI coding agents
Plain explanations and practical how-tos for building software with AI agents, whichever tools you use.
Basics
What coding agents are, the tools they run in, and the standards they share.
Glossary AI coding glossary: plain definitions of the terms you meet when building software with AI agents, from ACP and AGENTS.md to prompt caching, RAG and worktrees.
AI coding AI coding explained: the three kinds of tools (autocomplete, agents, app builders), which fits which job, what it costs, where it goes wrong, how to start.
What is an AI coding agent? What is an AI coding agent? A program that reads your code, runs commands and edits files in a loop until the task is done. How it works, and where it goes wrong.
What is an AI harness? What an AI harness is, what's inside one, an example loop in 20 lines, why the same model scores 14 points apart in two harnesses, and which harnesses exist.
What is a harness? What is an AI coding harness? The program around the model: loop, tools, context, permissions. How it differs from the model, and why it changes the results.
Terminal coding agents Terminal coding agents explained: how Claude Code, Codex CLI and OpenCode work in your shell, why developers pick them over IDE agents, and when not to.
What is MCP? What is MCP? The Model Context Protocol in plain words: what it does, how it differs from an API, how to add a server to your coding agent, and when to skip it.
What is a TUI? What is a TUI? A full-screen, keyboard-driven app inside the terminal. How TUIs work, why AI coding agents use them, the main frameworks, and a small Go example.
Agent Client Protocol The Agent Client Protocol (ACP) lets coding agents such as Claude, Codex or OpenCode run inside Zed, JetBrains IDEs and more. How it works and how to set it up.
AGENTS.md explained What is AGENTS.md? A Markdown file that tells coding agents how to build and test your repo. A short example, what to include, and how Claude Code and Codex read it.
Tutorials
Step-by-step guides you can follow at your keyboard.
Release notes with AI Generate release notes with AI: pipe git log into Claude Code, Codex CLI or OpenCode, get grouped Markdown back, and run it in CI. Commands, prompt and the traps.
Coding agents on Windows Which AI coding agent runs natively on Windows and which needs WSL: install commands, shell, sandbox and config path for each, plus the Windows traps that bite.
Codex on Windows Codex for Windows, step by step: the desktop app, the CLI from PowerShell, npm or WSL, signing in, the Windows sandbox and its errors, and fixes for crashes.
Claude Code on Windows Install Claude Code on Windows 11 from PowerShell, CMD or WinGet, pick Git Bash or PowerShell, decide on WSL, and fix PATH, proxy and duplicate-install errors.
OpenCode with local models Set up an OpenCode local model: the exact opencode.json for Ollama, LM Studio, llama.cpp and vLLM, the context-size fix, tool calls, and Claude Code or Codex too.
Working with agents
Day-to-day habits: writing tasks, using git, reviewing what the agent wrote.
Git with AI agents A git workflow with AI agents: start clean, one branch per task, commit checkpoints, undo safely, and run agents in parallel with worktrees. Commands to copy.
AI-generated code, measured AI generated code, measured: two Go codebases by one author, hand-written in 2025 and AI-assisted in 2026, with line counts, test ratios and what I changed.
Review AI-written code How to review AI-written code: read the diff, not the summary, then check scope, tests, invented APIs and error handling. Git commands and a checklist to copy.
Write good agent tasks How to write good tasks for coding agents: say what's wrong, what done looks like and what not to touch. A reusable template, before-and-after examples, sizing tips.
Models and cost
Choosing a provider and a model, running models locally, and what it all costs.
Choose a coding model How to choose a model for coding agents: what matters more than leaderboards, how to compare models on five tasks from your own repo, and when to mix models.
Coding benchmarks Every AI coding agent leaderboard worth checking in 2026: what each measures, who runs it, why the scores mislead, and how to benchmark agents on your own repo.
Which AI provider? Which AI provider for coding agents: subscription or API key, cloud account or local, and which providers work in Claude Code, Codex CLI and OpenCode.
Local LLMs for coding How to run local LLMs for coding: which model fits your GPU, the Ollama setup, the context setting that breaks agents, and connecting Claude Code, Codex or OpenCode.
Best Ollama models for coding The best Ollama model for coding at 8, 12, 16, 24 and 48+ GB VRAM and on Ollama Cloud: exact tags, tool support, context memory and the benchmarks behind them.
Best Ollama harness The best Ollama harness for your GPU: how Claude Code, Codex, OpenCode, Aider, Goose, Cline and DeepSeek Harness handle local models, setup effort, and what to pick.
Qwen Coding Plan The Qwen Coding Plan is closed to new Lite buyers and Alibaba now points to Token Plan. What each gives you, which agents work, the limits in words, and setup.
Best Qwen model for coding The best Qwen model for coding by job: qwen3.8-max, flash and plus in the cloud, and exact Ollama tags for 12, 24, 48 and 64 GB of memory, checked November 2026.
AI coding costs AI coding costs explained: why one agent session can send hundreds of millions of tokens, what sessions really cost, and the habits that cut the bill.
Subscription comparison AI coding subscription comparison: ChatGPT, Claude, Copilot, Cursor, Google, GLM, Kimi, MiniMax and OpenCode Go, the tools each plan works in and how limits work.
Codex: API key vs plan Codex API key vs subscription: what changes between ChatGPT sign-in and an OpenAI key, the exact login commands, how to check usage, and how to compare the cost.
Under the hood
How the machinery works: embeddings, RAG, caching and memory.
A terminal client, three times A terminal client for AI agents must show the work and ask inline, but never own the rules. What building beam three times taught, and why a terminal is for testing.
Prompt caching Prompt caching explained for AI agents: how it works at Anthropic, OpenAI, Gemini and locally, what silently breaks it in agent sessions, and how to keep it warm.
What is the KV cache? What is the KV cache, and why does a long context need so much GPU memory? The formula, worked examples for real coding models, a calculator, and how to shrink it.
What are embeddings? What are embeddings? Lists of numbers that capture what a text means, so similar meanings land close together. How they work, a short example, and the traps.
What is RAG? What is RAG? Find the passages that answer a question, then let a model answer from them. A 50-line example, the traps, and when long context or search is better.
What is a vector database? What is a vector database? A store that finds the embeddings closest to a query, fast. How it works, a pgvector example, the traps, and why you may not need one.
Agent platform for teams AI agent platform architecture for teams, from a real build: shared files with access rules, live shared chats, approvals anyone can answer, per-member billing.
Cloud AI platforms
Managed services from Google, AWS, Microsoft and OpenAI for building AI apps, under all the names they have had.
GenAI app builders GenAI app builders explained: what Google, AWS, Microsoft and OpenAI offer, which builders were retired in 2026, and when to write code instead.
Google Agent Search Agent Search was Gen App Builder, Vertex AI Search and Conversation, Agent Builder, AI Applications and Vertex AI Search. What it does today and how to use it.
AI IDE: built, killed An AI IDE is an editor with an agent inside. I built one on Eclipse Theia, killed it after three months, and here is what I use instead and why.
Bedrock Agents Classic Amazon Bedrock Agents is now Bedrock Agents Classic, closed to new customers since 30 July 2026. Who keeps access, what still works, and how to move to AgentCore.
OpenAI Agent Builder shutdown The OpenAI Agent Builder shutdown is on 30 November 2026. What stops working, how to export workflows to the Agents SDK, and where Evals users should go.
Foundry workflows retirement The Foundry workflows retirement ends declarative workflows on 1 December 2026. What stops, what keeps running, how to export your YAML, and where to migrate.
Safety and control
Permissions, sandboxes and secrets: keeping an agent inside its lane.
Agent permissions AI agent permissions for coding agents: a safe starting rule set, real config for Claude Code and Codex, the gaps in command rules, and when to skip prompts.
Keeps asking for permission Claude Code, Codex, Copilot or Cursor keeps asking for permission? The real cause in each tool, the exact rule or setting that fixes it, and what to keep asking.
YOLO mode YOLO mode turns off a coding agent's approval prompts. The exact flag in Claude Code, Codex, Copilot, Cursor and OpenCode, what each one turns off, and what's left.
--dangerously-skip-permissions What Claude Code's --dangerously-skip-permissions turns off, what still applies, the Codex, Copilot, Cursor and OpenCode equivalents, and safer options.
Sandbox AI agents How to sandbox an AI coding agent so a mistake, a prompt injection or a poisoned package can't reach your keys: built-in sandboxes, containers, and the traps.
Keep secrets from agents How to keep secrets away from AI agents: block .env reads in Claude Code, OpenCode and Codex, stop keys leaking through commands, and what to do if one leaked.
Approval workflow: the inbox An AI agent approval workflow for unattended agents: every gated call becomes one durable row, answered once from anywhere. What we built, broke and learned.