SpeakCode Blog — Notes on Agentic Development

• Claude Code vs Cowork vs OpenClaw: Which Surface for Which Job — A coding agent, a knowledge-work agent from the same company, and a self-hosted assistant. Why comparing their feature lists is the wrong move, and what actually decides it.
• What Claude Code Actually Costs When You Run Agents in Parallel — Official pricing answers what a plan costs, not what happens when you run four agents. Why burn scales linearly, where tokens actually go, and the levers that genuinely help.
• CLAUDE.md, AGENTS.md, and What Agents Should Remember — Claude Code does not read AGENTS.md — one import line fixes that. Then: where instruction files load from, why splitting into imports saves no context, and what belongs in a hook instead.
• Claude Code Permissions and Sandboxing: How Much to Actually Let It Do — Permission rules constrain what Claude decides; the sandbox constrains what the OS allows. Why that gap matters, what yolo mode still refuses, and where the real risk sits.
• Autonomous Claude Code: Auto Mode, Loops, and What Actually Runs Unattended — Auto mode is a classifier, not a yes-to-everything switch. The three mechanisms called loops, what survives unattended, and the rules from running 42 of them.
• Claude Code Plugins and the Marketplace: How They Differ From Skills — A plugin is not a different thing from a skill — it is a shareable, versioned wrapper around one. What can go inside, how the two marketplaces differ, and why to read hooks.json first.
• Running Claude Code Remotely: Web, SSH, and From Your Phone — The DIY answer is now mostly the wrong answer. What ships first-party, the three cases where self-hosting still wins, and why echo prediction over a network is a dead end.
• Claude Code Agent Teams: Running Parallel Agents That Don't Collide — Two agents in one checkout will destroy each other's work within the hour. The worktree layout, the claim field, and the merge gate that stop it — with the exact commands.
• Claude Code Desktop App Options on macOS — There is now an official Claude Code desktop app, and it does most of what people installed third-party GUIs for. The five surfaces, what each is best at, and the three gaps that survive.
• Claude Code Skills: What They Are and When to Write One — A skill is a playbook with a trigger attached, and the trigger is one sentence of frontmatter. What that means in practice, and the rule we use to decide when a procedure earns its own file.
• Claude Code Tutorial: The 20-Minute Setup Most Guides Skip — Most Claude Code tutorials stop at the install command. The step that changes outcomes is the 20 minutes you spend on CLAUDE.md and permission modes before the first real task — here's the full path to a multi-file change that lands.
• Claude Code Workflows That Actually Ship Code (vs. Chat With Extra Steps) — We run 42 autonomous Claude Code loops that each open one small, verified PR per tick. The lesson: workflow structure, not prompt quality, is what ships code — small increments, worktree isolation per task, and a merge gate that curls the real page.
• Codex vs Claude Code: An Honest 2026 Comparison (We Run Both Daily) — We run Claude Code and Codex side by side every day: Claude Code drives multi-file agentic work through its CLAUDE.md/hooks/skills ecosystem, and Codex adversarially reviews every diff before merge. The honest verdict isn't either/or — it's orchestration.