Agentic coding tools, Claude Code, Codex, and the agent modes in GitHub Copilot, work across a whole repository on multi-step tasks rather than completing the line you are typing. Used well they change cycle time, test coverage and how much of an engineer’s day goes on the work only an engineer can do. Used badly they produce confident code that looks like yours and is not. The difference is setup, and it is the same setup whichever tool you choose. This is ours.
Project instructions first
Every repository carries an instructions file the tool reads before it does anything: what the project is, how it is structured, the conventions, the libraries in use and the ones that are banned, how tests are run, what “done” means. It is the single highest-leverage artefact in the setup, because it is the difference between code that looks like the team wrote it and code that looks like the internet wrote it. Keep it short, keep it current, and treat a wrong answer from the tool as a bug in the instructions first.
This website is built that way: an instructions file at the root, a content model the tool must respect, a rule that colours and fonts live in one file, and a build-and-check step before any commit.
Standards the tool can follow
Naming, structure, error handling, logging, the test framework, the formatter. Written down, in the instructions or linked from them, in a form the tool can apply rather than a form a human would nod at. Where the team has a linter and a formatter, the tool runs them; where it has an architecture decision record, the tool reads it.
Review gates proportionate to risk
AI-generated code merges on the same terms as any other code, and the terms are explicit: tests pass, static analysis passes, secrets and dependencies are scanned, and a human reviews in proportion to risk. A documentation change needs a glance; a change to authentication needs a careful read by someone who owns it. The tool can draft the pull request and explain its changes; it does not approve them.
Where it pays back first
Test generation, unit and integration, against code that already exists. Test-coverage optimisation. End-to-end tests with a framework such as Playwright. Refactors with a clear specification. Documentation that has drifted from the code. Issue triage and first-pass investigation. These are the tasks where the tool is strongest and the risk is lowest, and they are where a team should start.
Budgets and measurement
Usage limits that match how the team works, so nobody is throttled mid-task and nobody is surprised by the bill; cost visibility by project; and a baseline taken before the rollout, cycle time, review load, defect rate, so that in thirty days the team can say what changed rather than how it felt.
Security
Enterprise terms that keep the code out of training data. Rules about what may be pasted where. Secrets kept out of prompts by design, which usually means out of the repository too. Dependency and licence scanning on anything the tool adds.
Claude Code, Copilot or Codex
They suit different work. Copilot is strongest as in-editor completion across a wide range of languages. Claude Code and Codex are agentic, working across the repository on multi-step tasks. Many teams run one of each. We decide on the team’s codebase and stack, not in the abstract, and we take no margin on any of them.
The setup is delivered under AI developer tooling, and the enablement that goes with it under training and enablement.