Categories
AI agents for Software: 28 real use cases
Shipping code, reviewing PRs, migrations and fixing things at 3am. 28 real use cases. 21 were done with 8 named agents, most often Claude Code, OpenAI Codex and Claude; 7 with custom-built or unnamed ones. Each links to its source.
28 use cases · 9 agents · Updated Oct 16, 2026
All 28 Software use cases
-
Run a team of 28 coding agents to rebuild an app designer
GoodBarber's CTO ran 28 coding agents, up to 20 at once, to restructure an app designer's code; he reports 569 pieces handled in six days, with 0 gaps across 12,074 test cases.
-
Build an Android app in 28 days with parallel Codex runs
OpenAI engineers had several Codex sessions write most of the Sora Android app in 28 days; OpenAI reports it reached #1 in the Play Store on launch day.
-
Port a Python tool to TypeScript with hundreds of agents
Anthropic's Mike Krieger had hundreds of Claude Code agents rewrite a Python tool as 165,000 lines of TypeScript over a weekend, Anthropic reports, checked against the original.
-
Rebuild old game code from games you own with agents
Developer i12bp8 built decomp.wtf, where coding agents turn retro game code back into readable code; the guide shows one Nintendo 64 piece matching exactly, with no overall results.
-
Upgrade old web projects to a new framework version
A developer had Codex write a reusable set of upgrade instructions for the Angular web framework; he reports it upgraded three open-source projects to version 22 within a week.
-
Build a Mac app that stays awake with the lid shut
A non-developer had Factory's Droid build LidAwake, a free app keeping his Mac awake with the lid shut; he says it took about three hours, including a launch video.
-
Trace and question a rogue group of web-scraping agents
The swarmcha.se team lured a group of data-scraping agents with hidden text and questioned them through a Claude agent; they say one reported running in a Tencent test environment.
-
Hand Devin a month of real coding and research work
The Answer.AI team gave Devin 20 real tasks over a month; they report 14 failures, 3 successes and 3 unclear results, with no way to predict which would work.
-
Move your website off a content system into plain code
Cursor's Lee Robinson used agents to move cursor.com off its website-editing system; the event page says it took 3 days and $260, with no downtime.
-
Have agents write, run and fix their own code
PwC rebuilt its software work with CrewAI agents that write code, run it and fix their own errors; a PwC leader says accuracy rose from roughly 10% to 70%+.
-
Test whether code comments help Claude Code build features
Roberto Aranda had Claude Code build two features three times each, with and without 168,000 comment lines; he reports the quality scores were too close to matter.
-
Hide passwords and customer data from your coding agent
A developer built an add-on that hides passwords and personal data before Claude Code sees them; he reports tasks still get done while real values stay on his computer.
-
Build and publish a simple web tool from one prompt
A developer gave Cursor a detailed spec written with Claude's help; it built and published a dollar-and-peso tip calculator, which they say went live in eight minutes.
-
Turn real-world failures into automatic tests
Rox's team had Devin turn records of their AI product's real failures into automatic tests; they report test coverage rising from 43% to 72%.
-
Let an agent debug chip designs in a live Verdi session
A developer built a connector letting an AI agent run Synopsys Verdi, a chip-debugging program, to trace design bugs; he reports testing it on four designs.
-
Teach coding agents reusable skills from passing tests
A developer built MuscleMemory to save reusable skills from coding-agent sessions once tests pass; they report Codex used 30.4% fewer tool calls and 25.5% less input.
-
Start system-problem investigations from a chat command
PostNL's AI team built a Slack command that has AWS DevOps Agent investigate system problems, so on-call engineers spend less time gathering evidence while keeping the decisions.
-
Check coding agents' claims against their real changes
Daniel Nwaneri built claimcheck to test whether coding agents' descriptions match their actual code changes; he reports it caught all 11 dishonest agents across 5 runs of 28, flagging no honest ones.
-
Turn an app's code into a priced cloud setup plan
Developer bealesh built a Claude Code add-on that turns app code into tested cloud setup files with price estimates; they report sample apps costing $77 to about $801 monthly.
-
Let client and agency coding agents share one app safely
An agency owner built CommonForge, a free, openly shared tool that lets client and agency coding agents work on one app safely; no usage results are given.
-
Turn a design file into a photography portfolio site
A developer had Claude Code read exact sizes, fonts, colours and photo crops from a Figma design to build a portfolio site; he directed the animation himself.
-
Have an agent test and fix its own browser add-on work
A developer had Claude Code test, load and screenshot his Chrome add-on after each change and fix failures itself; he estimates about three in four debugging moments are gone.
-
Work out a game's rules and build your own copy
A developer used a tool called REA with Claude Code to work out how Wordle scores guesses, including repeated letters, then build a working copy shown in a video.
-
Rewind a failing test to hunt down a software bug
Antithesis let Claude use its rewind tool to replay a failing test and probe a bug in HashiCorp software; the company shows a replay but reports no final fix.
-
Have an agent fix a bug in core system code
Linus Torvalds let Claude Code write a fix for part of the Linux system core; The New Stack reports reviewers then found a timing bug in it.
-
Share one Obsidian vault as memory for 8 agents
The author uses one Obsidian notes vault as shared memory for eight AI agents, including Claude Code and Codex, with a nightly agent fixing outdated facts.
-
Build a screen-sharing app with AI writing the code
Former CoScreen founders had AI models write much of Petal, a free screen-sharing app for coding together; they report under eight weeks, versus almost six years for CoScreen.
-
Train a simulated fly brain to review code changes
Nick Nisi trained Google's simulated 176,422-neuron fruit fly brain to judge code changes; he reports it agreed with automatic tests on 290 of 301 changes it had never seen.
