How software gets built with AI.

The pyramid

Six layers, ordered by how much of the engineering you take on, from website builders at the base to AI-native engineering at the top. Higher is not better. It means more of the work is yours, and more of what ships is yours to maintain. Find the layer you work at, then read the one above it to see what the step would ask of you.

6. Layer 6 AI-native engineering

   Building software that calls models while it runs, rather than using models to write it. A separate discipline, with its own costs in evaluation, security and spend.

   • AI-native software engineering

5. Layer 5 Orchestration

   The rung where one agent stops being enough and the work becomes coordinating several agents or tools across a single change, with you holding the plan.

   • Multi-agent software engineering
   • AI-orchestrated development

4. Layer 4 Delegated execution

   Handing an agent a whole task and reviewing what comes back. It pays when the task carries a check that proves it is done, and costs you when it does not.

   • Autonomous coding agents

3. Layer 3 Agent in the editor

   Working in a codebase you own with an agent beside you in the editor or the terminal. You stay the author, and every line you accept is yours to explain.

   • AI-assisted traditional coding
   • AI IDEs
   • Terminal and CLI agents
   • Design to code

2. Layer 2 Prompt to app

   Describing an application and getting a working one, with sign-in, data and a deploy. Fast to start, and the rung where the step to owning the code is steepest.

   • AI app builders
   • AI backend and database builders

1. Layer 1 Prompt to site

   Describing a website and getting one back. What prompt-to-site tools are good for, where their ceiling sits, and what a generated site still leaves you to own.

   • AI website builders
   • Screenshot to app

Cross-cutting practices

Testing, review, running what you build, cloud environments, voice and cloning. Six practices that apply at every layer of the pyramid rather than sitting on one.

• Cloud development environments

  The editor, terminal and build on a machine somewhere else, defined by a file in the repository. What it frees a constrained machine from, and the connection, bill and trust it asks for instead.

• AI testing and debugging

  Generating tests, reading stack traces, reproducing a failure and proposing a fix. Why a generated suite can confirm the bug it should catch, and what keeps tests answerable to the requirements.

• AI DevOps and cloud engineering

  Containers, pipelines, infrastructure and monitoring, written with a model. Where a generated mistake becomes an outage, a leak or a bill rather than a red test, and how to review it before it runs.

• AI code review, refactoring and maintenance

  Models on code that already exists: first-pass review, mapping an inherited codebase, refactors, upgrades and the documentation nobody wrote. Why a confident first pass can stand in for the judgement it skipped.

• Voice and conversational coding

  Speaking the work instead of typing it, to a transcript or to an agent that acts. Strong for plans, specifications and hands that cannot type all day, and every spoken word still needs editing as text.

• Reverse engineering and cloning

  Recreating an existing product from its screens or its observed behaviour. Why this section does not recommend it, why there is no course on it, and where the legitimate half of the skill is taught instead.

Reference

• Glossary

  The terms these pages use for building software with AI, from rung and brief to stand-in, seam and prompt injection, each defined as this section means it and linked to the page that covers it.

• Tool index

  Products by name for each of the eighteen categories: who makes each one, what it is, a link to its own page and the day it was last checked, with renamed and replaced tools kept under their old names.