Coding is the first place AI stopped being an assistant. These tools are why the price of custom software has fallen far enough that this site sells a Software Factory built on them. Even if nobody in your firm will ever open a code editor, the category is worth ten minutes, because it changes what you can afford to commission.
The three categories at a glance:
| Category | The developer’s role | Leading examples | What a buyer should know |
|---|---|---|---|
| In-editor assistants | Drives every keystroke | GitHub Copilot | Table stakes for any supplier |
| Agentic tools | Directs and reviews | Claude Code, Cursor | Where the cost drop comes from |
| App builders | Describes and curates | Fast-moving field | Superb for prototypes, unsafe for production |
In-editor assistants (GitHub Copilot)
An in-editor assistant watches a developer type, completes the next lines, answers questions and explains unfamiliar code. GitHub Copilot defined the category and is still the workplace default, built into the most common developer tooling, and every major provider’s models now power similar features across editors. The gains are real but bounded, because the developer still drives every step. For a firm buying software, a supplier using assistants is table stakes, not a selling point.
Agentic coding tools (Claude Code, Cursor)
Agentic coding tools are given tasks rather than keystrokes: read this codebase, implement this feature, run the tests, fix what fails, and show me the result. The developer becomes a director who reviews and approves rather than types. Claude Code is the strongest of these in my daily use, and you can check that claim: it is how the systems on this site’s Labs page get built, inside the specification-and-gates method the factory page describes. Cursor brings the same approach into an AI-first editor, and the frontier labs all now ship agent modes of their own. This category is why build timelines have shrunk so far; the agentic AI cluster explains the supervision that keeps it safe.
AI app builders
App builders turn plain-English descriptions into running web applications, with the interface, database and logic scaffolded in minutes. The products change too fast for a durable list, but the category is stable, and its best business use is narrower than the demos suggest. As production systems, builder output tends to fail the unglamorous tests: security, data protection, edge cases, upkeep. As prototypes they are superb. A working mock of the tool you want, built in an afternoon, is the best specification a commissioning firm has ever been able to hand a builder. Prototype cheaply, then build properly.
What a non-technical firm should take from this
The short version: prices have dropped, so re-ask old questions, and judge suppliers on their review discipline rather than their tools.
- Re-price the shelf. The system you assumed you couldn’t afford may no longer be out of reach, and the build you shelved years ago deserves a fresh quote.
- Change the supplier question. Ask candidates how they use agentic tools and, more importantly, what review, testing and security gates sit around them, because the tools amplify whatever discipline exists.
- Read the risk honestly. AI-built code without human review ships mistakes faster too. Method, not tooling, is what you are buying.
The guide to choosing an AI consultancy gives you the questions to put to a supplier.