How software gets built now

AI now writes code, and small teams ship in weeks. The work around the code changed too.

01Where the work slows down now

For a long time, delivery was limited by how hard it was to build and release safely. Cloud, managed services and AI have cut much of that friction.

The harder questions now are whether the code should exist and whether it fits the architecture. Your team also has to check it fast enough without storing up problems.

Many teams now produce more code than they can check.

02How software used to be built

Not long ago the default flow was heavier by design: requirements, planning, design, implementation, QA, release and maintenance, with handoffs and approvals to manage risk.

That model existed for good reasons. Infrastructure was harder to run, automation was weaker and late changes were expensive.

It produced a lot of strong software. It also made teams slow to experiment.

03What cloud, platforms and AI changed

Cloud made infrastructure quick to set up, and managed services and modern frameworks meant fewer layers built from scratch.

Then AI made code itself cheaper to produce. Teams now use it to scaffold features, write tests, draft documentation and speed up refactors.

Google Cloud’s 2025 DORA report frames successful AI adoption as mainly a systems problem. Three other studies add numbers.

• CNCF’s 2024 survey reports 89% of organizations using cloud-native approaches to at least some extent.
• In Stack Overflow’s 2025 survey, about 70% of agent users agree agents cut the time spent on specific tasks and 69% that they raised productivity. Only 17% say agents improved team collaboration.
• Microsoft Research found that developers using GitHub Copilot finished a representative coding task around 55% faster.

04The bottleneck moved to trust

Trusting new code means checking that it works, fits the architecture, is secure, can be maintained and solves the right problem.

The teams that do well review code as fast as they generate it, and release without losing control of the system. The amount of code generated is a poor measure of progress.

05Shorter cycles, more review before release

Strong teams now work in shorter cycles, with more testing, security scanning and review before release.

OpenAI’s Codex guidance says agents do better in tidy repositories, with clear docs, reliable tests, a configured environment and written standards in files such as AGENTS.md.

• Clean repositories decide how much AI can help.
• Testing and security checks move earlier and matter more.

06What this means for each role

For founders, a bigger team is no longer the automatic answer. Small teams can ship a lot, as long as they don’t generate more than they can check.

For technical leads, the job shifts toward designing how work gets delivered. Standards, reliable tests, security scanning and clear ownership decide whether AI speed helps or adds debt.

For developers, the value moves from typing every line to judging what AI writes and keeping the architecture sound.

07Five steps for building with AI

When code is cheap to generate, a badly framed problem costs more.

1. Frame the problem before you generate output.
2. Keep repo setup, commands, tests and ownership clear and written down.
3. Use AI for scaffolding, tests, docs, migrations and repetitive coding work.
4. Review for correctness, security, architecture fit and product relevance. Nothing merges until the tests, the security checks and a person sign off.
5. Ship small changes, so a mistake affects as little as possible.

08Where this is going

Software teams will likely keep getting smaller and lean more on cloud services and AI. The best of them will also be more disciplined than many teams were before.

People still decide the architecture and sign off on security.