Agentic coding workspaces

What it is

An agentic coding workspace is a tool that works inside a folder rather than inside a chat window. Claude Code, Codex and Cursor are the common examples. You give it a task in plain language; it reads the files it needs, proposes edits, runs commands, checks the result and reports back. Despite the name, the folder does not have to contain code. A folder of Markdown notes, a spreadsheet export, a set of email drafts or a grant application all work.

The difference from a chat box is not intelligence. It is that nothing needs pasting in or out. The tool sees the real files, and its output lands in the real files.

Why it matters for an EA organisation

Most of the operational work in a small organisation is not writing; it is moving information between places. Reformatting a report, reconciling two lists, turning meeting notes into actions, renaming forty files, updating twelve pages when a policy changes. A chat box helps with each of those one at a time and leaves the moving to you. An agent does the moving. That is where a session stops saving ten minutes and starts saving an afternoon.

It is also the tool class that most needs two things this guide series keeps returning to: a context layer, so the agent knows the organisation, and clear permissions, so it cannot do harm quickly.

The moves that give most of the value

1. Start in a folder that is safe to change and easy to restore. A copy, or a Git repository where every edit can be undone.
2. Read the context file first. If there is no CLAUDE.md or AGENTS.md, write four sentences before the first task: what this folder is, what to never touch, how to check a change worked, and the house style.
3. Give tasks with a checkable end. Not “tidy the notes” but “every note gets a date line at the top in this format; show me the list of files changed”.
4. Ask for a plan before a large change, and read it. Thirty seconds of reading saves an hour of undoing.
5. Let it run the check itself. If there is a test, a build, a word count or a link checker, tell the agent to run it and report the output. Do not take “done” on trust.

How to do it this afternoon

1. Pick one small, real, annoying task. Renaming files to a convention, adding front matter to notes, turning a long email thread into a decision list.
2. Copy the folder, or confirm it is in Git and the working tree is clean.
3. Install one tool. Claude Code runs in a terminal; Cursor is an editor; Codex runs in a terminal or the ChatGPT app. Any one is fine. Choose the one someone nearby already uses.
4. Open the folder, write the four-sentence context file, and give the task with its checkable end.
5. Read the plan. Approve it. Watch what it does the first time; after that, you will know what to trust.
6. Check the result yourself against the checkable end, then look at the diff or the changed files once more.
7. Add one sentence to the context file about anything it got wrong.

What good looks like

• A task that was three hours of clicking is done in fifteen minutes, and the person who owned it checked the result and agrees.
• The agent asks fewer questions each week because the context file answers them.
• Nobody is afraid of it, because everything it did is visible in a diff and can be undone.
• People give it tasks in the morning and read the results after lunch, instead of watching it work.

Common mistakes

• Running it in the live folder with no backup and no version control. The first time it deletes the wrong thing, adoption ends.
• Turning off the permission prompts on day one because they are annoying. They are annoying for a reason. Loosen them for one folder at a time once trust exists.
• Vague tasks. An agent given “make this better” makes it different, not better.
• Treating it as a coding tool only. The finance lead reconciling two exports gets more from it than most developers do.
• Never reading the plan. The plan is where you catch the wrong assumption before it becomes forty wrong edits.

When not to bother

A single document that needs a better paragraph does not need an agent; the chat box is faster. An agent earns its place when the task touches many files, needs a command run, or has to be repeated. If the organisation has no shared folder, no Git and no one who has used a terminal, start with a context file in a chat tool’s project feature and come back to this in a month.