Claude Code or Codex: which coding agent should an engineering team use?

Use both, and choose per task: Claude Code suits interactive work in your own repository where you steer as it goes, and Codex suits well-specified tasks you can hand off and review later, though each can now do some of what the other is known for. The choice matters less than the team practices around it: tests as the contract, a review of every diff, written project instructions and deliberate permissions. Employers are already naming both. Of 416 software engineering postings we collected on 29 September 2026, 27 named Claude Code and 12 named Codex.

This piece compares working styles, not benchmarks or prices. Both products change often, so check the current documentation before you standardise on a feature.

What is the difference between Claude Code and Codex?

Both are coding agents: you describe a task, the agent reads the code, edits files, runs commands and tests, and shows you what it changed. The difference is where each one's centre of gravity sits.

Claude Code is Anthropic's coding agent, and its home is the terminal inside your repository. You work alongside it: it proposes a plan, asks permission before risky actions, and you correct it mid-task. A team shapes it with a CLAUDE.md file of project instructions checked into the repo, hooks that run your own scripts at set points (a formatter after every edit, a block on touching a protected folder), skills that package a repeatable procedure, subagents that take a side task in their own context, and MCP servers that connect it to tools such as an issue tracker or a database.

Codex is OpenAI's coding agent. It runs in the terminal too, but it is best known for delegation: you hand it a task, it works in a sandboxed copy of the code in the background, and it comes back with a diff to review. Several tasks can run at once. A team shapes it with an AGENTS.md file of project instructions in the repo, and with approval and sandbox settings that decide what it may do without asking.

The table is a tendency, not a wall. Both vendors keep adding the other's strengths.

Which one do employers ask for?

Both, and usually alongside other tools. Our sample is the same public fetch behind our count of AI job ads, re-run on 29 September: Hacker News "Who is hiring?" threads for July to September 2026, Remotive and Arbeitnow. Among 416 software engineering postings, 67 (16%) named an AI coding tool or asked for AI-assisted development. Claude Code appeared in 27 (6% of all 416), Codex in 12 (3%), Cursor in 24 and GitHub Copilot in 10. Most ads that name one tool name several and leave the choice to you, which is the more useful signal: employers want fluency with agents, not loyalty to a brand.

The small print: these counts are small, the sources lean towards startups and Europe, Australia barely appears, and some postings are duplicated. Treat the gap between 27 and 12 as a snapshot, not a market share. Do software engineering jobs expect AI coding tools? covers the full count, and the data is public.

Which tasks suit which tool?

Choose by how well you can specify the task and how much you need to watch it.

Reach for an interactive session (Claude Code's default) when:

• You are exploring a codebase you did not write and do not yet know what the change should be.
• The task crosses many files and needs judgement calls along the way, such as a refactor or a migration plan.
• You want to stop the agent the moment it heads the wrong way.

Reach for a delegated task (Codex's default) when:

• The task is small, clear and testable: a bug with a reproduction, a dependency upgrade, a test to add.
• You have several such tasks and would rather review five diffs than type through five sessions.
• A failing test already defines done, so you can judge the result without watching the work.

Do neither, or check by hand, when the change touches authentication, payments, secrets or data migrations and you cannot verify the result yourself.

A practical split for a team: interactive sessions for design and hard changes, delegation for the queue of small well-defined work, and the same review standard for both.

What practices should a team adopt whichever tool it picks?

Four, and they matter more than the tool:

1. Tests as the contract. A person writes or approves the test first; the agent's job is to make it pass without editing the assertion. Correctness is defined by the team, not the model.
2. Review every diff. Agent-written code goes through the same pull request review as anyone's, and the author can explain every change in their own words. A diff nobody understands does not merge.
3. Project instructions in the repo. Keep CLAUDE.md and AGENTS.md short, specific and version-controlled: how to run the tests, the conventions, the folders not to touch. Many teams keep one source and point the other file at it, so the two agents are told the same thing.
4. Deliberate permissions. Decide what the agent may run without asking, keep production credentials out of its reach, and use sandboxes or worktrees for parallel work. A hook or approval rule that stops a bad change beats relying on everyone remembering.

Should a team standardise on one tool?

Standardise the practices, not necessarily the tool. One shared tool keeps onboarding, instructions and hooks simple, a good default for a small team. But the ads we counted mostly list several tools, and engineers who work in both styles are worth more than engineers who know one product's menus. Pick a default, write the instructions so the other agent can use them, and let people delegate small tasks to whichever fits.

Where should an engineering team start this week?

• Write a one-page CLAUDE.md or AGENTS.md for your main repository: test command, conventions, no-go areas.
• Pick three small tasks with failing tests already written. Run one interactively, delegate two, and compare the review effort.
• Agree the review rule in writing: every agent diff is read and explained by the person who merges it.
• Decide what the agent may run without approval, and put production secrets out of reach.
• After two weeks, look at what came back from review and adjust which tasks you delegate.

Square 1's Claude Code for Engineering Bootcamp is twelve weeks, live on Zoom with one instructor, about 15 hours a week, for engineers who already write software in any language. Its six blocks each end in a deployed project and a gate: a feature shipped through Claude Code with a review log, a legacy refactor with tests added first, a team workflow of hooks, skills and a review gate, a hardening pass, a feature from spec to production defended in a recorded viva, and an employer brief with a hiring sprint. The Codex for Software Engineering Bootcamp follows the same structure around specs, delegated parallel work and test-first fixes. For a shorter start, Claude Code from Zero and Codex from Zero are on-demand courses recorded by an instructor and graded by Nova, the AI tutor, and AI Coding Workflows builds one feature in Cursor, Claude Code and Codex and teaches when to reach for which. All are on a waitlist today.