The agent loop: read, plan, act, verify

From a suggested line to a finished task

GitHub Copilot began in 2021 by finishing the line you were typing. An agent is a different kind of tool: given a task, it reads the files and commands it needs, decides what to do, makes the change itself, and checks whether the change worked. GitHub Copilot itself now ships this pattern too, as agent mode in the editor and a Copilot CLI with a Plan mode and an Autopilot mode, alongside Cursor’s Agent mode and Background Agents. This course focuses on the two tools most teams reach for first, Claude Code and Codex CLI, and uses them for every worked example.

Anthropic’s Claude Code and OpenAI’s Codex CLI both started as terminal-native agents, and both now run inside VS Code through an official extension, so the same read-plan-act-verify loop is available whether you work from the command line or the editor. Claude Code added subagents, skills (both covered later in this course) and cloud sessions you can start and return to later; Codex CLI added multi-step plans, repository-stored skills and subagents you can use to split up a larger investigation, plus cloud tasks with the same hand-off-and-check-later pattern.

The same four steps, in Claude Code and Codex CLI

Read means the agent inspects the relevant files, tests, and sometimes a running process, before proposing anything. Plan means it states an approach, ideally one you can approve or correct before code changes exist. Act means it edits files or runs commands inside whatever the harness allows. Verify means it checks its own work against tests, a build, or another form of evidence, rather than declaring success by assertion.

A worked example in Claude Code: prompted with “The signup form accepts an email with no @ character. Find why, fix it, and tell me how you checked,” it read src/forms/signup.ts and its test file first, stated a one-line plan (add a guard clause before the existing regex check), made the edit, then ran the test suite. Its final message: “Added a guard clause for missing @ in validateEmail(). Ran npm test — 47 passed, 0 failed.” Four steps, one sentence each, no vendor jargon required to follow any of them.

• Claude Code: reads via its own file and search tools, offers a plan mode you approve before anything changes, edits files and runs commands inside a permissioned session, and can eliminate routine prompts with its classifier-based auto mode once you trust the pattern; verification runs through your tests or hooks.
• Codex CLI: multi-step plans, skills stored in the repository, and multiple concurrent subagents for splitting up a larger investigation; sandbox modes from read-only through workspace-write to full access pair with an approval policy that decides whether it asks first, and cloud tasks let you hand a run off and check on it later (as of mid-2026).
• GitHub Copilot’s agent mode and CLI, and Cursor’s Agent mode and Background Agents, run the identical loop under their own menu names and their own approval settings.

Why the harness decides how much to trust it

The harness is the program wrapped around the model: what it can read, which commands it can run without asking, whether it can reach the network, and where a human has to approve before anything proceeds. A capable model inside a permissive harness with no verification step is not more productive than a careful one; it is simply further from view before something goes wrong.

This is why Claude Code separates what a session may do automatically from what needs your explicit approval, and why Codex CLI pairs a sandbox mode (what it can technically touch) with an approval policy (when it must ask first) rather than treating autonomy as one on/off switch. Cursor’s Background Agents and Copilot CLI’s Autopilot mode raise the same question under their own settings. The vendor names change every year; asking what this agent can read, what it can change without asking, and how you will know if it got something wrong does not.