90 real AI agent use cases in engineering, each with its source and a prompt to try it yourself.
After $45 in GitHub Actions overage in one month, a solo founder put an idle homelab to work as CI runners, with Claude making it easy to set up. Daily spend dropped from $2-4 a day to under $0.25.
Result Even runs some jobs faster than GitHub's own workers
A first homemade synth has 36 hall-effect keys and buttons and 7 rotary encoders. The firmware started by hand, then gentle prodding at Claude got everything needed much faster. Rev 2 of the PCB is nearly done.
Result Hope to one day sell a few
Markdown files should just open, so they asked Claude Code for a viewer. It opens instantly: faster, they say, than any other macOS app, and faster than Bitwarden's Chrome extension shows its fill menu.
Result Came out of the same codebase as their JSON viewer
Migrating a password reset flow, Claude Code and Opus produced code on par with a mid level developer. It needed close supervision throughout, and cost roughly 2 months of Pro subscription for about half a day's work.
Result A bit slower overall than doing it by hand
Working on a Quake project, Claude decided on its own to build a test suite: demo playbacks with snapshots compared pixel by pixel, plus recorded demos whose screenshots it checked for each element.
Result Used at most 1% of a $20 plan's weekly limit
A URL shortener side project ran on AWS. Claude Code wrote the GitHub Actions workflows for the move and fixed each failure as it came up. The writer reports zero downtime.
Result Estimated hosting: about $44 a month, down from $120
A 7-app monorepo needed an 8 GB memory setting and 25 seconds to typecheck. Claude Code made the changes for the move to oxlint, oxfmt and tsgo, and the team validated each step.
Result No more 8 GB memory setting in CI
A data scientist tried Cursor on a real project. Prototyping, code completion and commenting went well, but pandas code kept failing and it deleted a bunch of data files.
Result Code completion and commenting both graded A
Building their first AWS project by vibe coding, the writer designed the architecture and let Codex write the code. A flawed first design took a day, against an estimated three or four; the redesign also took a day and scaled.
Result ChatGPT's thinking mode caught a concurrency bug
After building a minimal proof assistant, the writer asked Claude how they knew the prover was actually sound. Claude checked, found five different ways it could prove false theorems, and fixed them.
Result Also proposed a fuzzer that turned out useless
Four years after shelving a homemade RSS reader and selling off its $500 domain, the writer built it this year with Codex and Claude. They use it daily, along with 30 other strangers who found it public.
Result 30 strangers now use it too
Chasing a childhood itch for the exact 1996 After Dark Flying Toasters screensaver, the writer turned to Claude. It did a lot of good reverse-engineering work, but kept inventing details that needed pushback.
Result Released as AGPL on GitHub
During a live call, the writer dumped nomad-job-inspect data for a 10,000-node cluster into Claude Cowork and asked for memory/CPU requirements by job, datacenter and SKU. It built the table in under 30 seconds.
Result Estimated at 3-4 hours done by hand
A rabbitmq connector occasionally got stuck after rare network drops, too minor to spend days fixing. Claude built a test harness that reproduced it, and the actual fix wasn't too hard.
Result Automatically runnable scenarios, not a one-off check
A client's CRM needed a new deals feature. Devin opened a first pull request in under 10 minutes, but its PRs carried unneeded changes, it gave a false answer about a library, and it hit its usage limit.
Result Said a library wasn't used elsewhere when it actually was
A 42k-line site with packages to upgrade: 19 easy, 42 medium and 14 hard, by Cursor's own sort. Cursor cleared the easy and medium ones, then a background agent moved it from Remix v2 to React Router v7.
Result Merging agent PRs is "starting to feel routine"
Adding Oracle to a database migration tool took 9 months and about $160K. A pipeline of Claude coding agents, run like a teaching hospital, added IBM Db2 in less than two days for $4172 in tokens.
Result No human reviewed the Db2 code yet; that check is underway
20 real engineering tasks over a month, from a Notion-to-Sheets sync to debugging, handed to Devin, the autonomous AI engineer. It closed with 3 successes, 14 failures and 3 inconclusive results.
Result It once spent over a day chasing an impossible task
Six AI workers, two each on Claude, Codex and Gemini, were set loose to build a SQLite-like engine in Rust. It passed 64 queries of a 1,000+ query suite, with slop like redundant buffer copies and no concurrency.
Result 54.5% of all commits were just lock coordination
Turning plan files into Claude Code workflows took 5-6 hours each. A script that converts them deterministically, and flags what can move out of the LLM, cut that to 20-30 minutes at about 20% of the tokens.
Result Also removed a 10-15 minute workflow construction wait
Their Claude Max200 plan started including API credit. In 2 minutes Claude built a fallback callout moving their busiest app stage from Gemini on OpenRouter to Haiku. Their OpenRouter bill should drop about $200 a month.
Result OpenRouter stays on as overflow and for testing models
An inherited app on Azure, an error, and a hunt that could easily take 1-2 days by hand. With CLI access to Azure and the repo, Claude Code ran the config and network checks and found the root cause in about 45 minutes.
A MongoDB-like database for a hobby project, written spec-first by Claude, Kiro, Qwen Coder and Cursor. They read zero lines of the code and checked it through tests. The first version took about two weeks, part time.
Result Learned more in two months than in 20 years
Running Claude Code in YOLO mode risked projects leaking tokens, so a container setup isolated each project's permissions. Typing a hotkey and a feature name, like text-to-speech streaming, got it done in five minutes.
Result From nothing on screen to words at every pause
A Linux box mattered but wasn't their day job, and fixing it when it broke was anxiety-inducing. They gave Grok Bot root SSH access and their old setup notes, and now run nginx restarts and firewall rules from a phone app.
Result Reachable from a phone app to fix problems anywhere
In SpaceXAI's own guide, engineer Lingxi Li's bots start Cursor cloud agents, check every PR every 30 minutes and merge the low-risk ones. A teammate shipped 2,000+ PRs in a month, they report.
Result Low-risk PRs merge themselves
Their own code, their own database, on their own laptop. OpenAI's models and Claude found the issues but refused to confirm them. Grok tested that they were exploitable and suggested a fix.
Result Tested on their own copy, fix proposed
lldb behaved differently through Python in Docker than on their own Linux box. Grok 4 traced it to how address sanitizer works in each environment. They're impressed so far.
Result Traced to address sanitizer differences
A messy, roughly 900-line Blender script, abandoned. Cursor turned it into a proper repo for all their 3D designs, then built shapes they described in plain language. "These aren't 1-prompt creations", they add.
Result Now a mono repo for all their 3D designs
A huge, convoluted codebase they'd taken over. One prompt in Cursor's Plan mode found the bugs and planned the fix, then the agent made it, where the client's previous team hadn't.
Result Plan mode found every bug in one prompt
They planned the architecture with Gemini first (dependency injection, UI components, Firebase, logging), then let its agent mode build it step by step. The app works like a charm.
Result Works like a charm, neatly structured
The malware was cleaned out, but a hidden backup in the database, triggered by the WordPress cron, would have brought it back. Gemini pulled the encrypted packages apart and spotted it.
Result Traced to a fake Cloudflare popup scheme
Buildkite wants GitHub webhooks for every repo, the kind of setup that's tediously annoying by hand. Codex built ephemeral runners on refurbished Mac minis and migrated each repo. Payback: two to three months.
Result Their colleagues are now considering it too
Claude proposed the changes, opened a Jira issue with small subtasks, and shipped them as reviewable PRs with tests. A coworker had guessed two weeks; Claude and ChatGPT guessed 6 to 8 weeks.
Result Shipped as reviewable PRs with tests
Assembled boards were too expensive for this KiCad 10 project. Claude found cheaper versions of their parts, then reworked the schematic for through-hole parts they'll solder themselves.
Result Ordered bare boards; parts from Mouser and Amazon
They'd written one by hand years before, so they knew exactly what they wanted. In VS Code's Agent Mode, Gemini built it and refactored it into a Python decorator. One dumb corner case slipped through.
Result One dumb corner case slipped through
Code from 1999, around 100 hours of one developer's time, and every new model failed. Sonnet found nothing. Opus 4.5 managed a workaround. Opus 4.6 solved the whole thing.
Result Sonnet found nothing; Opus 4.5 only a workaround
A very subtle but important bug in open-source lattice Boltzmann code. Claude didn't give up until it found it. Later they realised they could have just had Claude write a whole new solver.
Result Back when Claude was still cheap to run
It started as a hand-drawn layout for primes by a graphic designer. Gemini and ChatGPT turned it into C++, and the ~3.1x speedup came from cache use, not new math.
Result The speedup came from cache, not new math
Detailed Markdown specs go to OpenClaw, which spawns and tests Claude Code instances for hours until the app works as described. "Is the code any good? Probably not." They check before trusting it.
Result Sometimes builds for 2+ hours unattended
Reviewing every turn kills the value. Working from a spec and a to-do list, Claude Code runs for weeks on a compiler and a strategy game while they step in now and then.
Result A spec and a to-do list keep it on track
Stuck on a Swift error, they paste it with the surrounding code and ask what's going on. ChatGPT usually gives a solid lead in about half a minute. They still review every line.
Result They still review every line it writes
Making Vestiges playable without a mouse took nearly two days of part-time work, with Claude busy throughout. Then they set up an MCP server so Claude can play the game like a player would.
Result Nearly two days of part-time work
Ask about the page you're reading and the chat already has its text. Claude Code built the Streamlit app in a single file, correctly, on the first attempt.
Result Page text added to every question automatically
They already run local reviews with Codex and Claude Code. Assigned as a reviewer in GitHub, Copilot finishes in a couple of minutes and raises valid issues they'd never found.
Result Review done within a couple of minutes
They wanted game audio through the controller's headphone jack while streaming to an Android TV box. With the spec, a USB capture and live access, Claude built it, for $20 instead of begging a developer.
Result Game audio through the controller's jack
Every visitor picks a World Cup winner, and countries light up by majority vote. When the live dot layer choked on mobile, they rewrote it in WebGL to keep it smooth.
Result Seeded with 4,200 predictions across 183 nations
Their Hermes agent on Qwen took a C++ entropy extractor from 12MB/s to about 48, then kept optimizing alone. Six hours later it stopped, just past 100MB/s, with a correct implementation.
Result Stopped on its own after 6 hours, correct
Karatsuba multiplication and a recursive halfgcd, from academic papers into an existing codebase. A Qwen-powered Hermes agent took many tries, but its own tests caught the errors.
Result Its own mutation-tested tests caught the errors
Their Hermes coding agent works unattended for about 15 minutes, opens a PR and pings them to review. At basic coding tasks, they say it's faster than they are.
Result Faster than them at basic coding tasks
They pasted raw PPP frames in hex into Gemini, which decoded and explained every byte. An hour later they knew enough to fix a Zephyr OS driver bug and submit a patch.
Result Every byte decoded and explained
Never built for mobile, a parent of three, a busy life. Gemini guided the Kotlin architecture and the UI, and their first Android app was a real joy to build. Impossible a couple of years ago, they say.
Result A real joy to build, for a parent of three
They worked with Claude Code like an engineering team: requirements, design specs, test plans, step by step for about 3 months. The app scores QR codes before the phone acts on them.
Result Started as just for them; now polished
Their agent wakes Claude at night to work through the to-do list. Having watched them struggle exporting Android icons, it wrote a resizer that outputs every size and folder, unprompted.
Result Another night: a Go tracker for their test matrix
By hand it would easily have taken a day. Claude Code cleared them while they watched TV, and has since added a feature to a Rust editor they don't even know.
Years of Oh My Zsh for just 3-4 features, and a slow shell to show for it. Claude Code rebuilt those features into a single .zshrc, and the shell got much faster. Their other configs moved in minutes.
Result Moved to Lazy.nvim and Ghostty in minutes too
Remote desktop into an existing Linux session never worked quite right. Claude rewrote xrdp: connect, pick display :0 or :1, sign in, done. No more fiddling with x11vnc.
Result Drop in, pick a display, sign in with PAM
Give it detailed tasks, agree the approach, then let it iterate for hours. Their 20 years of experience steps in only for obscure bugs, and sometimes they play with the dog while it works.
Result Now working like a tech lead, not a coder
Which bike races are on TV today, and where? The data is messy and the sites break every week. Each day Claude reads the unmatched-race report and proposes fixes, so one person can keep it running.
Result Legal TV listings for every race, every day
Their reTerminal E1002 e-ink display needed a dashboard and firmware that accepted all 6 colours. Claude did both, after earlier reverse-engineering their NAS's LEDs. They vibe-debug 10x more than they vibe-code.
Result A Things replacement on the side, still janky
They wanted big local database imports to show progress and a finish time, and to stop the machine sleeping. Copilot built exactly that, plus an MCP server so it writes better queries.
Result Plus an MCP server for better queries
OpenClaw checks Sentry and the product to-do list daily and opens PRs for what it can fix well. Within 2-3 days it was handling a lot of their manual work, even when they're away from the desk.
Result They keep coding even away from the desk
Micromanaging five, 10 or 20 agent sessions doesn't scale. Their OpenClaw agent, Patch, runs the Claude Code instances in a shared tmux session and reports back in one Telegram chat.
Result One Telegram chat instead of 20 sessions
Half their effort goes to automated adversarial review. Claude plans and writes the code, Sol tears into it with only the coding standards to go on, and it costs about $300 per developer a month.
Result Fewer bugs and clearer documentation
An undocumented legacy codebase mixing two languages, a SCADA platform and proprietary hardware logic. With little help, Claude Code worked out most of it, with only a couple of misunderstandings.
Result Ready for the refactor the team had planned
The secret is all in the spec. They pressure-test a structured one in Kiro for hours or days, then hand it to Kiro CLI and Claude Code in tmux. Usually the agents finish without stopping.
Result A one-line TDD rule lifted every test's quality
Thousands of words of planning on their phone over the holidays, then Claude Code built it. It forgot one key feature until asked, and their son's reaction was one of the best looks they've had.
Result Built with a test suite and a clear to-do list
They knew every detail of the old code. Cursor rewrote it in 2 days with less technical debt, they happily archived the old repo, and the experience pushed them to retrain toward AI evals.
Result Old repo happily archived
They kept running into usage caps. Claude built a status line showing usage, reset time and context, plus a hook that catches spikes, like a long chat suddenly resending everything.
Result Spikes caught before they burn the quota
An always-on GPU box costs far more than cheap storage. Claude wrote a small Python script that boots one on Linode from a persistent drive in about 2 minutes, so they pay only while it runs.
Result Boots in about 2 minutes, paid only while on
Some of those bugs had beaten a 30-year veteran for over a week. Claude Code ran a /loop around the clock on a prototype compiler, fixing over 500 crashing test cases while they glanced in now and then.
Result Including bugs that beat a 30-year veteran
Flakes, Sway, Home Manager, secrets, even vim, all asked for in prose. Claude Code set up the old ThinkPad, and when an update blacked out the screen, it found the open bug and rolled Gnome back.
Result Zero regrets, even after a black-screen bug
The game doesn't show all your cats at once, and you can have 100+. Within a day Claude Code built a PyQt6 tracker; two months on, it has 600 users and Chinese, Russian and Polish versions.
Result Translated into Chinese, Russian and Polish
Authenticate, pick the cluster, find the namespace, run kubectl, every single time. Claude Code built a tool that remembers it all: "a huge time saver" they'd never have had time to build.
Result A huge time saver they'd never have built
Instead of plugins, the agent writes down improvement ideas, edits its own Rust source and recompiles itself. It runs on DeepSeek 4 Flash on two local GPUs at about 138 tokens a second.
Result About 138 tokens a second on local GPUs
They needed tooling to spawn a Go proxy and run integration tests, using only Perl's standard library. Claude Code wrote most of it, and it beat their usual shell scripts for stability.
Result More stable than their usual shell scripts
Hooks firing at once on the same event clashed: two wrote the same state file and corrupted it. Running each event's hooks in turn from cached input fixed it, for about 200ms per prompt.
Result About 200ms overhead per prompt
On a codebase of 500k+ lines, they hand separate tasks to several agents in Antigravity and review only the 10-20% that matters, in minutes. No meetings, no friction.
Result Reviews take minutes, on 500k+ lines
Frontend race conditions hide from you. Cursor's debug mode came up with several theories, built a harness to test each one, cleaned up after itself and handed back a fix.
Result Cleaned up its own debugging code
Code from one AI is hard to trust blind. This developer runs a Docker harness where Claude writes and Codex reviews, works in it most of the day, and open-sourced it as moarcode.
Result Open-sourced as moarcode
The script came out almost perfect on the first try, except for a one-character typo in the file-path code. Finding it took an hour: the hidden cost of code that's nearly right.
Result Close to really good, frustrating when wrong
Two ancient V100 GPUs, one open-source inference engine, and about a week of Claude optimizing it. Decode speed reached around 300 tokens a second, and the work is still going.
Result Still going after about a week
No source code, just disassemblers. Told it had Ghidra and Radare2, Claude took the game apart, ported it to Rust and Bevy, and added MIDI music without being asked.
Result Added MIDI music without being asked
A 4-year-old React anatomy app needed to reach more of the world. Claude Code internationalized it for 45 major languages in one session, covering ~85-90% of the world's population.
Unicode characters were rendering wrong in PDFs. Claude Code raced through several approaches in an afternoon to find each one's limits, ending in a custom font fallback. Then a day of polish.
The auth prototype came together fast, until the last step: it sent the user ID as a plain string next to the token. Caught in review and fixed, it still beat starting from scratch.
Result Caught in review, and still faster than scratch
Optimizing a Taichi kernel had each agent looping on its own. Putting Claude Code and Codex in one discussion to combine their ideas cracked it, saving several hours.
Salesforce wanted faster pull requests without losing quality. After an org-wide Cursor rollout, PR velocity rose 30%+ and legacy-test time fell 85%, according to Cursor.
Engineering managers needed more QA, reliability and DevOps capacity than their teams had. Litera added Devin agents in those roles, and test coverage rose 40%, according to Cognition.
Result Test coverage up 40%
New data integrations used to need days of engineering. Stepstone rolled n8n out company-wide and now runs 700+ workflows, parsing millions of job documents a month, per n8n.