Coding‑agent primer video

A coding agent is software that works inside a project like a junior teammate: it reads files, changes code, runs commands, and comes back with a result to review. A new four-hour primer video, posted April 14, teaches OpenAI’s Codex in exactly those terms instead of as a prompt box for one-off snippets. (mycoding.id) (developers.openai.com) OpenAI’s own documentation describes Codex as an “AI coding agent” that can pair in a terminal, integrated development environment, or app, and can also be delegated work in the cloud. The help center says it starts from “a prompt or spec,” then navigates a repository, edits files, runs commands, and executes tests. (help.openai.com) The video outline tracks that workflow step by step, from “Agentic Coding vs. Coding Harness” at 13:39 to “Understanding the Agentic Loop” at 15:40, then later into plan mode, approvals, sandbox security, project rules, and sub-agents. Its chapter list also includes “Project Guidance via agents.md” at 1:22:05 and “Approval Policies” at 1:53:56. (mycoding.id) That emphasis matches the product OpenAI is shipping in 2026. The company says Codex can run multiple agents in parallel across projects in the Codex app, review pull requests in GitHub, and carry out tasks in isolated sandboxes that users can later review, merge, or pull down locally. (help.openai.com) The practical question in that setup is not only what the model can write, but what it is allowed to touch. OpenAI’s command-line interface documentation says Codex can read, change, and run code on a machine in the selected directory, and the video spends more than 10 minutes on approval modes, network access, and permission overrides. (help.openai.com) (mycoding.id) The primer also treats context as a resource that has to be managed. Its chapter list includes “Managing the Context Window,” “Truncation and Hallucination Issues,” and session controls such as “/new,” “/resume,” and “/fork,” which are the mechanics that keep a long coding task from losing track of the codebase. (mycoding.id) OpenAI’s developer materials push the same idea from another angle: Codex is presented as a system for “build and ship” work, not just autocomplete. The official learn page links out to guides on code review, codebase modernization, editor integrations, and a “Prompting Guide” for its recommended agentic coding model. (developers.openai.com) The command-line tool adds more of the workflow layer. OpenAI says the terminal client supports local code review by a separate Codex agent, subagents for parallel work, cloud tasks, scripting with an exec command, and even web search for up-to-date information needed to finish a task. (help.openai.com) Access has also widened beyond early power users. OpenAI’s help center says Codex is included with ChatGPT Plus, Pro, Business, and Enterprise or Education plans, and “for a limited time” is also included with Free and Go plans, while the Codex app is available on macOS and Windows. (help.openai.com) The through line in the new primer is simple: the hard part of coding agents is not getting one answer from a model, but setting up a loop of planning, editing, running, checking, and approving changes. OpenAI’s own docs now describe Codex in that same loop, which is why a tutorial on files, sessions, sandboxes, and human sign-off reads less like prompt engineering and more like operating instructions. (mycoding.id) (help.openai.com)