Should our staff build software with AI agents?

What AI coding agents are

An AI agent is an AI that can take steps on its own towards a goal, not just answer a question (Field Guide 17). A coding agent does this for software. As their makers describe them:

• Claude Code (Anthropic) is described as an agentic coding tool that reads your codebase, edits files, runs commands and works with your development tools, in a terminal, code editor, desktop app or browser.
• OpenAI Codex is listed across the ChatGPT apps, a command-line tool, an editor extension and the cloud, for tasks such as understanding code, building features and fixing failures.
• Google Antigravity is described as an agentic development platform with an app for managing several agents, a command-line tool and a full editor (Field Guide 39).

In practice, someone describes what they want in plain language and the agent writes, runs and adjusts the code. People often call this vibe coding.

Which small tool would save your team an hour a week if someone could build it safely this month? Book a free AI Value Briefing and we will look at it with you.

What non-developers can safely build

The UK National Cyber Security Centre (NCSC) puts it simply: different code deserves different levels of oversight. It names prototypes, demos, internal tools with limited exposure and projects with no sensitive data or security functions as places where lighter oversight is reasonable. Good first builds for staff include:

• a tidy-up script for a spreadsheet export;
• a small calculator or checklist page for a team;
• a prototype that shows IT what a process could look like.

Field Guides 14 and 15 explain how to describe what you want clearly before the agent starts.

Where IT must be involved

The same NCSC guidance lists areas that need tight control: login and access logic, sensitive personal data, secrets such as passwords and keys, safety-critical systems, and anything where a security flaw would have significant consequences. In plain terms, bring IT in before a tool:

• reads or stores customer, staff or health data;
• handles passwords, access keys or payments;
• connects to live business systems or can change records;
• is used by people outside the team that built it, or by customers.

Often the right answer is that the staff prototype becomes the specification, and IT or a developer builds the production version.

Review and security

The NCSC notes that AI models learn from large amounts of existing code, some of which has security flaws. The agent makers say the same about responsibility: Anthropic's Claude Code security page states that users are responsible for reviewing proposed code and commands before approving them, and that no system is completely immune to attacks such as prompt injection.

• Keep the agent's permissions narrow: one project folder, no admin rights, approval before commands that change things (Field Guide 26).
• Never paste real passwords or keys into a prompt or into the code.
• For anything beyond a personal helper, have a second person review it, and use your normal security checks.
• Use organisation-managed settings so IT can set the limits once for everyone.

Ownership of what is built

A tool that one person built on a laptop becomes a problem when that person goes on holiday or leaves. Agree simple rules:

• Every tool has a named owner and a short note of what it does and what data it uses.
• Code lives in an approved company location, not a personal account.
• Tools that others rely on are listed in a simple register that IT can see.

Who owns the rights in AI-assisted work is a separate legal question; our article Who owns what AI creates? covers it.

Goes deeper in our free AI Field Guide: 14. Building small apps with AI, without coding, 15. Specs: write it down before AI builds it, 17. What an AI agent is (and is not), 26. Guardrails and permissions: what AI may and may not do, 39. Google Antigravity: building software with agents.