Skip to content
anythingai.space
Frameworks & Toolsintermediate13 min · 9 steps

Coding Agents

How Claude Code and its peers work: the agent loop and tools, CLAUDE.md, subagents, permission modes, safe CI pipelines, cost, and the wider coding-agent landscape.

TWO WAYS TO WORKAutocompleteyou typeyou typeyou typeCoding agentreadplanedittests failfixtests passPRtime →The real difference now: where the loop runs, and who setsits permissions
requesttokentool callresponse
1 / 9
ProblemStep 1 of 9

A coding task is a loop, not a completion

Autocomplete predicts the next few lines while you type. A real task (“add auth to this API”) needs reading the codebase, planning, editing several files, running tests, fixing failures and only stopping when tests pass.

Every major tool now has an agent mode, so the useful question is no longer suggest vs implement. It's where the loop runs and who controls it: a laptop, CI, a vendor cloud, or a self-hosted runner.

Lesson 5 of 6 in Frameworks & Tools