You've used AI chat tools — maybe you already run a personal assistant like OpenClaw or Hermes on a server. These guides teach the mental models behind real coding agents like Claude Code and Codex: the terminal, servers, sessions, git, context, and the discipline that keeps agent work safe. Every guide is interactive — simulations you can click, not screenshots.
What a coding agent actually is, how it differs from the assistant you already run, and how to pick your first one.
An LLM in a loop with tools, operating on real state. Chat answers questions; agents produce outcomes.
Read → 02You already believe in agents on servers. What changes when the agent touches code — and the security lessons of 2026.
Read → 03The real decision axes, why many builders run both, and how to run a fair one-week bake-off on your own repo.
Read →Agents need a place to live. The terminal, a server, a way in, and sessions that survive you closing the laptop.
Terminal vs shell vs process, pipes, exit codes — and why agents live in inspectable text.
Read → 05SIGHUP, controlling terminals, and the session-server model that makes agent work survive anything.
Read → 06What key pairs actually prove, the first-connection prompt, and hardening a box agents live on.
Read → 07What a VPS is, honest sizing for agent work, and the first ten minutes on a new box.
Read →The discipline layer: memory, specs, git, and running many agents at once without them destroying each other's work.
Tokens, context rot, compaction — and why the repo itself is your agent's only durable memory.
Read → 09Why vague prompts fail structurally, and how acceptance criteria turn "plausible" into "verified done".
Read → 10Content addressing, the Merkle DAG, snapshots and pointers — and the failure modes agents introduce.
Read → 11Worktrees, one-writer-per-branch, merge order — parallelism is free only when agents share nothing.
Read →Fresh from the pipeline — newest first.
Eighteen months of unreconciled accounts. Why balances come before categories, and the four ways a matcher lies to you.
Read → 13Running it on a timer with no model involved — durable stores, movements not balances, and what a 2× error means.
Read → 14Rules first, model second, humans on the edge — and the double-count that balances perfectly while doubling your costs.
Read → 15Why a script that fans out to disposable subagents beats a checklist loop — with a 253-agent image pipeline you can inspect.
Read → 16The same method as a 3-minute cheat sheet — the two primitives, the barrier that cost 91 minutes, and the grader that rejected 38 of 57 images.
Read → 17Two subscriptions are two rate-limit windows. How to dispatch a fan-out across both, read each meter for free, and grade work on the vendor that did not produce it.
Read →