AI agents for coding: 68 real examples
Agentic Feats lists 68 sourced examples of AI agents doing coding work; 42 link the artifact; most common agent: Claude Code (26). Newest first.
- Feats
- 68
- Shown
- 4262%
- Agents
- 13
- Updated
- Oct 13
Agents used for coding
counted from 68 feats
| Agent | Feats | Shown | Newest |
|---|---|---|---|
| Claude Code | 26 | 18 | |
| Devin | 12 | 3 | |
| Codex | 9 | 9 | |
| Replit Agent | 9 | 7 | |
| Cursor | 8 | 4 | |
| Other | 4 | 2 | |
| Custom agent | 3 | 3 | |
| Grok Bot | 3 | 1 | |
| Hermes Agent | 2 | 2 | |
| OpenClaw | 2 | 0 | |
| Claude in Chrome | 1 | 0 | |
| Gemini agent | 1 | 1 | |
| Perplexity Comet | 1 | 1 |
A live blog aggregator pulling posts from Dev.to, AWS Community Builders, Bravo LT and RedMonk, built and deployed in a single day.
Blog aggregator in one day
Claude Code built and deployed a Vue 3 blog aggregation site from a year-old requirements document, rewritten as a focused prompt.
A live MVP website for the podcast and newsletter went up after about 10 hours of struggling to load it on a custom URL.
Podcast website built with Replit
The author used Replit to generate a clickable website prototype for their podcast and newsletter, then finished the MVP with help from Val Town.
Within one night Devin produced a working unanswered-emails interface using the Gmail API, after the author wrote the authentication code themselves.
Unanswered email tracker with Devin
Devin wrote a Gmail function to find emails needing a reply, built a front end, tested it locally, and deployed it to a public link.
The app went live on a custom domain, dashdiary.co.uk, for a total build cost of $35.82 plus a few dollars for the domain.
Road safety reporting web app
Replit Agent built a full web app where the public reports bad driving, including a Postgres database, image uploads, Firebase login, and a landing page.
The author shipped a working summariser targeting a cut from 2-3 weekly hours of newsletter reading to 30 minutes.
Newsletter summariser app
Replit Agent built a Gmail newsletter summariser for a low-coder, with batch summaries, deduplication, archiving and export, over three days.
Small scoped fixes took about five minutes to test and merge versus roughly an hour by hand; the full-stack task failed and was rewritten.
One month testing Devin
Devin was connected to the RunLLM GitHub repo and given small bug fixes and one larger full-stack task over a month.
After 2-3 review rounds the PR was production-ready and merged. The task took about 1 hour versus roughly 15 minutes coding it alone.
Third-party token API endpoint
Devin AI planned and built a backend API endpoint for generating third-party user access tokens, then opened a GitHub pull request.
The extension was accepted and published on the Chrome Web Store, and the author says it solves their own habit of looking down at notes.
Teleprompter Chrome extension published
Replit Agent built a floating teleprompter overlay extension for browser video meetings, refined with Claude and ChatGPT.
A roughly 3,000-line open-source Canvas tool built in four months. It cut a 45-minute weekly student check to a 2-minute conversation with Claude.
Canvas LMS tool via MCP
Claude Code co-developed a Canvas LMS MCP server with an educator who is not a developer, covering courses, grading, analytics, and student data anonymization.
The agent finished the page in about 5 minutes; the author reviewed the PR, tested locally, and merged it to production.
Pricing page via background agent
Cursor's background agent built a pricing landing page from a short spec, working on its own branch and opening a pull request.
The author, a non-developer for a decade, built a working end-to-end booking app with artist dashboards in 48 hours. It is not yet launched.
Tattoo booking app prototype
Replit Agent built a booking and payment app for tattoo artists from prompts, covering calendar, Google OAuth, Stripe, and Postmark emails, over one weekend.
One prompt produced the thumbnail section with accurate colours, layout and button hover states, and it beat the earlier Lovable attempt built from static images.
Figma design to HTML
Cursor, connected to Figma's Dev Mode MCP, turned a selected Figma design into HTML and CSS from one prompt containing the design URL.
Get five like these every week
The week's strongest feats, each with its receipt, plus a short roundup of AI news. Free.
You're on the list.
A free weekly newsletter. See what's in it
A production micro-SaaS went live in about 3 hours, versus an estimated 3 to 4 weeks solo.
Micro-SaaS built and deployed
The author used Replit Agent to build and deploy Calendar Autobot, an app that turns forwarded emails and images into Google Calendar events.
A working, deployed app that shows images, weather, and news for a city, finished in roughly 90 minutes from an empty repo.
City dashboard from blank repo
Devin took a blank GitHub repo and built a City Explorer dashboard with a FastAPI backend and React frontend, then deployed it.
It located all three bugs in one shot each, the Verify bug within minutes, saving the author hours of debugging.
Debugging a cryptography implementation
Claude Code debugged a fresh Go implementation of the ML-DSA signature algorithm, finding why Verify rejected valid signatures and two separate Sign bugs.
The author shipped 120 models, an MCP server for Claude Desktop, and 140 chaos tests with a 100% pass rate over 22 months.
Mental models framework built with agents
A solo founder gave Claude, ChatGPT, Windsurf, and Cursor defined roles to build and deploy a 120-model mental models framework.
About 3 hours and roughly $22 produced a live, searchable map of funded companies, versus an estimated 4 to 5 hours by hand.
Interactive map of funded companies
Replit Agent built and hosted a searchable map of funded brick-and-mortar companies on Wefunder from the author's prompts, using the Leaflet map library.
After the author added the suggested indexes, page load performance improved significantly in production.
Fixing a slow page with indexes
Cursor's agent listed likely causes of a production-only slowdown, traced NestJS database queries, and recommended which database indexes to add.
Comet generated multiple test payloads for different cases and produced a final report after the full test run.
API endpoint testing in Postman
Perplexity Comet ran API endpoint tests in Postman after development, working from payload context the author supplied.
Replacing one byte ($88 to $EA, DEY to NOP) kept the lives counter fixed at 3 when tested in an emulator.
Unlimited lives in River Raid
Claude, connected to Ghidra through an MCP server, analyzed an unlabeled Atari 8-bit ROM and found the code that decrements the player's lives.
The agents wrote over 1 million lines of code across 1,000 files, and a three-week Solid to React migration reached +266K/-193K edits.
Hundreds of agents build a browser
Cursor ran hundreds of concurrent coding agents, split into planners and workers, on one project for close to a week to build a web browser.
The build took about five hours in one evening and produced 627 passing tests. The author estimates the API cost at roughly $35.
Data analysis agent suite
Claude Code built a six-agent data analysis suite from a 713-line spec the author drafted with another AI.
Across nearly 2,000 sessions and about $20,000 in API costs, the agents produced a 100,000-line compiler that builds Linux 6.9 on x86, ARM, and RISC-V.
Parallel agents build a C compiler
A team of 16 parallel Claude Opus 4.6 agents wrote a Rust-based C compiler from scratch, coordinating through git locks with no orchestration agent.
The author debugged a custom WhatsApp integration hands-free while brushing their teeth, using the voice stack running locally on an M4 Pro.
Hands-free voice for Claude Code
The author built a local speech-to-text and text-to-speech stack with a custom MCP so Claude Code can be driven by voice through AirPods.
Four features finished in about 2 hours instead of 3-5 days each, and the daily view took 40 minutes.
Parallel feature shipping in Codex
Keith Rumjahn used the OpenAI Codex desktop app with git worktrees and parallel threads to build four features for their Mac app at once.
367 posts and 5 pages were migrated to MDX in one afternoon, about 4 hours, versus an estimated 3-5 days by hand.
WordPress blog to MDX
Claude Code, run as a terminal agent, pulled every post from the author's WordPress REST API and converted them to local MDX files.
Cognition merged 659 Devin PRs into its own codebase in one week, up from 154 in its best week of 2025.
Devin builds Devin
Devin handles Cognition's own engineering work across Slack, Linear, and the web app: PRs, code review, bug triage, debugging, and design system audits.
The author ran 4 to 8 agents in parallel and wrote 300+ feature specs in one project, then ported the system to new repos.
Parallel coding agents with specs
Claude Code agents run in separate tmux windows as planner, worker and PM, driven by numbered Markdown feature specs and six slash commands.
Cognition merged 659 Devin-authored PRs into its own codebase in one week, up from 154 in its best week of 2025.
Devin building Devin
Cognition's team uses Devin across Slack, Linear, web, CLI and API to write, review, triage and debug code in its own codebase.
The agent rebuilt readable Go source from the stripped binary, but got some route paths wrong, such as /my-messages instead of /me/messages.
Reverse engineering a Go binary
Cursor, connected to Ghidra through a GhidraMCP server, analyzed a stripped Golang HTTP binary and identified its libraries, API routes, and file structure.
The author shipped five projects in about two weeks, with public repos and a finance tracker running nightly on a Raspberry Pi.
Personal apps on a Raspberry Pi
Claude Code built five personal tools for a non-coder, including a multiplayer game, a WhatsApp scheduler, and a finance tracker.
On a dashboard filter task, Claude fixed URL query param handling in three iterations without the author touching the browser.
Local browser testing loop
Claude Code with the Chrome extension builds features, then opens localhost in the author's real Chrome session to test, read console errors, and fix issues.
All the author's old utilities, such as picture converters and sprite extractors, now compile and run in DOSBox. The main game does not compile yet.
Reviving a 1991 assembly game
Claude Code CLI with Sonnet helped the author recompile their 1991 DOS assembly code using UASM and run it in DOSBox.
Time to diagnose a CI failure fell from 45 minutes to 6, and test flakiness dropped from 8% to 0.8%.
CI failure triage agents
OpenClaw agents watched GitHub Actions failures, classified them, and posted analysis and issues, then grew into test generation, coverage scanning, and self-healing locators.
It found multiple remotely exploitable kernel bugs, including an NFS heap overflow undiscovered for 23 years. Five fixes or reports are linked in the post.
Linux kernel vulnerability hunting
Claude Code, run by Nicholas Carlini, scanned every file in the Linux kernel source for security bugs using a simple loop script.
The author consolidated 4 legacy repos into 2, moved MySQL to PostgreSQL, upgraded the full stack, and wrote full test suites, solo.
Solo product migration setup
Claude Code, configured with layered CLAUDE.md files, memory, and agent teams, planned and carried out a multi-repo product migration for one engineering manager.
Devin built a three-endpoint Express API in about 4 minutes, but chose the wrong pagination style on a real codebase and needed review and refactoring.
Devin one-week coding review
Devin was given a week of real coding tasks, from a simple Express API to pagination, and reviewed on speed and accuracy.
A legacy pipeline with a roughly 40% pass rate lost 20 minutes of runtime and gained 35% pass rate in weeks.
Fixing flaky CI tests overnight
Claude Code monitored 30 days of CI results, found the worst flaky tests, narrowed reproductions, and reran fixes overnight to verify them statistically.
Combined lint and typecheck time fell from 29.04s to 1.25s, about 23 times faster, with no memory config needed for typechecking.
Monorepo lint and typecheck migration
Claude Code migrated a 7-app, roughly 1,670-file TypeScript monorepo from Biome and tsc to oxlint, oxfmt and tsgo, fixing code to meet new rules.
About 12 hours produced 46 commits and over 14,000 lines of code across three shipped products.
Three products in one day
Replit Agent rebuilt a full-stack platform, added seven features to a live RAG app, and produced a demo video during a free Agent day.
The author shipped a live book-sharing app and its demo video, both made with Replit Agent, during a 24-hour buildathon.
GiveABook book sharing app
Replit Agent built GiveABook, an app where people list books they have read, drop them off, and neighbors pick them up on Sundays.
Test runs approved per day on Devin more than doubled in a couple of months, and engineers run 10 to 20 Devins in parallel.
Devin cloud testing of PRs
Devin writes a test plan, runs the app on a cloud VM with computer use, and returns annotated test videos and reports.
A prototype took two evenings and a live game six weeks; after seven weeks it had 164 players, 109 active in one day.
Telegram text RPG shipped
Claude Code wrote every line of Zapaleni, a Slavic folklore text RPG in a Telegram bot, for a project manager who does not code.
From one prompt it created a repo, wrote the app and tests, got CI green, provisioned Postgres and deployed behind HTTPS with no further messages.
Self-hosted agentic software factory
The author built a sandboxed home server stack where Hermes with Codex takes one prompt and runs the whole software lifecycle, from repo to deployment.
Claude gave failed builds for a week and used up the 20x Max weekly limit. Sol had worked 20+ hours without success.
VNC client build attempts
The author asked Claude, then ChatGPT 5.6 Sol, to build a bare-bones VNC viewer for Windows 10 that avoids dropped keystrokes.
Within 90 minutes Claude produced a functional proof of concept confirming an authentication bypass across four tested endpoints.
Reverse engineering a papercut zero-day
Claude Code analyzed a patch for an actively exploited Papercut NG vulnerability and worked to build a working exploit chain.
The best recorded run scored 39,359, reached 4,924 meters, and hit a combo of 117.
Browser game and PPO player
Grok built the browser game Roofline, then trained a PPO agent that plays the live page.
The fleet manages over 200 cloud agents at once, up from 15 managed by hand, and built Grok Bot iOS v0 in three weeks.
Engineer bots managing cloud agents
A SpaceXSI engineer runs five Grok Bot engineer bots that spin up Cursor cloud agents, review PRs via a Notion database, and auto-merge low-risk changes.
Non-test Python fell from 1,063,826 to 698,363 lines (34.4%), gateway/run.py from 34,847 to 5,512 lines, and functions over 300 lines from 192 to 2.
Million-line codebase refactor
Hermes Agent split a million-line Python repo into 36 groups and dispatched 1,393 subagents in git worktrees to refactor it.
Each model pass added about 1.5x to 2.0x speedup, reaching roughly 7.5x to 32x over baseline; one implementation beat xgboost on speed.
Faster Rust ML code via benchmarks
Coding agents wrote Rust ML algorithms like UMAP from scratch, then iteratively optimized them against criterion benchmarks.
No feats match this filter. Show all.
Proof marks:Shown artifact linked·Numbers a specific number stated·Claim claim only