Three agentic coding tools dominate the UK developer workflows we see day to day in 2026: Claude Code, Cursor, and OpenAI’s Codex CLI. They look similar from the outside — all three drive code with a frontier model — but they are optimised for very different jobs. Picking the wrong one for the work in front of you typically costs a fortnight before you realise.

Claude Code — long-running, terminal-native

Claude Code lives in your terminal. It runs sessions that can last hours and span dozens of files. It excels at tasks where the agent needs to keep a lot of context in its head at once: a multi-file refactor, a feature that touches both backend and frontend, an investigation across an unfamiliar codebase.

It is at its weakest when the task is small and crisp. Spinning up a Claude Code session to add a single optional argument to a function is overkill. It is also at its weakest in environments where you cannot give it broad filesystem access — locked-down corporate laptops, regulated industries with strict tooling controls. The right shape of work for Claude Code is ‘a feature that would have taken a senior engineer half a day.’

Cursor — IDE-resident, edit-in-place

Cursor lives where most engineers already live: an IDE that looks and feels like VS Code. Its strength is the immediacy of pulling an agent into whatever file you’re already in. Tab completion, inline edits, multi-file edits within a single conversation — all of it is friction-free.

Cursor is at its strongest when the engineer is steering the work continuously. The agent is a fast typist with good taste; the human is the architect. It is at its weakest when you want to delegate something for an hour and check back. The session model and context window are not designed for that. If you want a ‘go away and come back when it’s done’ workflow, Cursor is not the tool — Claude Code or Codex CLI is.

Codex CLI — scriptable, atomic, headless

OpenAI’s Codex CLI is the headless option. You hand it a task, it runs in a sandbox, it returns a diff. There is no interactive session in the same sense as Claude Code or Cursor. This makes it brilliant for two specific jobs: scripted bulk edits across many repos (think ‘update this dependency, run the tests, open the PR’), and building automation pipelines where AI is one step in a longer machine workflow.

Codex CLI is at its weakest when the task is exploratory or when the spec is incomplete. It does not push back on a vague prompt the way an interactive session does. Use it when you know exactly what you want and you want it done atomically, in many places, possibly in parallel. Do not use it as your primary day-to-day driver.

Which one for which job — a decision matrix

Quick spec, IDE-bound work, lots of human steering: Cursor. Multi-file feature work, exploration, ambiguous spec: Claude Code. Bulk operations, CI/CD integration, headless automation: Codex CLI. Mixed work? Use multiple. Many of the engineers we train run Cursor as their primary IDE, fire up Claude Code in a separate terminal for hour-long jobs, and reach for Codex CLI when they want to do something to fifty repos at once.

There is no ‘best’ tool here, just the right tool for the shape of the work. The teams we see making the most progress have all three installed and pick consciously. The teams we see stuck have one of them and try to make it do everything.

What we use for client work

Inside AI Project Fixers, our default is Claude Code for new builds and Cursor for live-editing existing client codebases. We use Codex CLI for repository-wide migrations and CI workflows. The choice is task-by-task and we coach our clients to do the same.

If your team is currently committed to a single tool and the velocity gain has plateaued, that is almost always the reason. We cover this in detail in our agentic AI coding courses — a half-day workshop is usually enough to retrain the muscle memory.