Case studies / Coding agents
Coding agents
Agents that write, review, migrate and test code, or build whole apps.
Approach
68 write-ups
Tag:
-
SnowflakeNewCoding agentsAgents Are Code Too: What I Learned Shipping a Cortex Agent Through Azure DevOps
Problem Deploying AI agents reliably across dev, test, and prod environments was error‑prone.
Solution They treated the Cortex Agent as code, storing its YAML spec in Git and using Azure DevOps pipelines to version, test, and release the agent together with its semantic view and search service.
Read on medium.comDetails -
JetBrainsNewCoding agentsJetBrains IDEs Transformed for the Agentic Era of Development – With Air
Problem Developers spend hours each week reviewing and fixing AI‑generated code.
Solution JetBrains added Air to its IDEs, letting developers run, track, and review coding agents inside the editor, with Git worktrees, notifications, and IDE tooling integration.
Read on blog.jetbrains.comDetails -
AtlassianNewCoding agents3 lessons from rebuilding Atlassian’s product development lifecycle for people and agents
Problem Engineers spent hours manually investigating and fixing security alerts.
Solution They created an agent pipeline that scans code, suggests fixes, runs tests, and only asks engineers to approve safe changes.
Read on atlassian.comDetails -
SourcegraphNewCoding agents
Keeping code moving with clearer review expectations
Problem Code reviews were slowing down development as agents generated more changes.
Solution They built an AI bot that examines pull‑request diffs, checks rules and author history, and automatically routes or approves routine changes.
Read on sourcegraph.comDetails -
ShopifyNewCoding agents
How we rebuilt the Shopify app in 12 weeks
Problem Developers needed to rebuild a huge mobile app quickly, but manual coding was too slow and error‑prone.
Solution They created a workflow where AI agents generate most of the Swift and Kotlin code, supported by clear architecture rules, tests, linting and CI feedback to keep the output correct.
Read on shopify.engineeringDetails -
PostmanNewCoding agentsHow to migrate JMeter tests to Postman
Problem Teams had to maintain separate functional and load testing tools and reports.
Solution Postman created a coding‑agent plugin that reads JMeter .jmx files, converts them to Postman collections, and generates performance run commands, letting both test types live together.
Read on blog.postman.comDetails -
PostmanNewCoding agents
Handling a leaked credential with the Postman plugin
Problem A leaked API token was discovered while the engineer was away from a computer.
Solution They created a coding agent with the Postman plugin that uses the CLI or MCP server to locate the leak, identify the source, and record an audit trail, all from a mobile app.
Read on blog.postman.comDetails -
SalesforceNewCoding agents
Using AI to Build AI: How Agentforce Engineers Scale Automation Without Sacrificing Reliability
Problem Engineers needed to automate building the Atlas Reasoning Engine without unreliable AI-generated code.
Solution They created an AI‑driven automation pipeline that generates and validates components, using tool calls and safety checks to keep reliability high.
Read on engineering.salesforce.comDetails -
PostmanNewCoding agents
Postman Plugin for Claude Code: A Hands-On Guide
Problem AI agents guessed API behavior, causing retries and wrong parameters.
Solution Created a Postman plugin for Claude Code that lets the agent run real Postman CLI commands for each API step, using skills, hooks, and gating to keep actions accurate.
Read on blog.postman.comDetails -
VercelNewCoding agents
Chat SDK agents can now stream generative UI with json-render
Problem Developers had to create separate UI code for each chat platform, slowing down agent responses.
Solution Created a JSON‑based UI spec that agents output, which json‑render turns into native cards for Slack, Teams, etc., streaming updates as the spec builds.
Read on vercel.comDetails -
MicrosoftCoding agents
Formally verifying agent-written full-stack web apps with PDD Kit
Problem Agent‑written code often drifts from the original specification, and tests miss many bugs.
Solution Created PDD Kit, a toolkit that lets agents generate formal proofs from natural‑language specs, ensuring the code matches the contract for all inputs.
Read on microsoft.comDetails -
ShopifyCoding agents
Reflex: an architecture for building mobile apps with agents
Problem Testing mobile screens required slow emulator builds, delaying agent-written code verification.
Solution Created Reflex, an MVVM architecture that runs app logic headlessly and exposes state via a CLI, letting agents test iOS and Android with a single test suite.
Read on shopify.engineeringDetails -
VercelNewCoding agents
Introducing Vercel Schedules
Problem Developers needed a way for agents and apps to run tasks on dynamic schedules without redeploying or managing separate databases.
Solution Created Vercel Schedules, an API/SDK that lets code create, update, or cancel one‑time and recurring jobs, which are triggered via queues while the app runs.
Read on vercel.comDetails -
VercelNewCoding agents
How Zapier turns ideas into working software in hours with Vercel
Problem Teams needed to show partners working prototypes quickly, but the process took weeks.
Solution They used Vercel’s v0 platform to spin up interactive, multi‑view demos in under two hours, letting backend engineers deliver clickable experiences without frontend design time.
Read on vercel.comDetails -
SentryNewCoding agents
Finding performance wins in our own traces with a coding agent
Problem Sentry needed to locate and fix performance bottlenecks in its own backend traces.
Solution They built a coding agent that used the Sentry MCP to analyze trace data, suggest code changes, and automatically remove hundreds of millions of unnecessary queries each week.
Hundreds of millions of queries removed weekly
Read on blog.sentry.ioDetails -
NVIDIANewCoding agents
Build Specialized Agents on a Stack You Control with NVIDIA NemoClaw
Problem Developers needed a secure, customizable stack to run specialized agents with enterprise data.
Solution NVIDIA released NemoClaw blueprints that bundle model, harness, and governance tools, letting teams plug into existing harnesses like OpenClaw, Hermes, or LangChain and deploy agents safely and efficiently.
Read on developer.nvidia.comDetails -
JetBrainsCoding agents
Getting Started with Air in JetBrains IDEs – Part 1
Problem Developers struggled to see what coding agents were doing and what still needed attention.
Solution JetBrains added Air to IDEs, integrating coding agents with a dashboard that shows task status, changes, and pending work.
Read on blog.jetbrains.comDetails -
ReplitNewCoding agents
Anyone can now bring their ideas into the company codebase
Problem Non‑technical team members can't directly modify core code, causing slow review bottlenecks.
Solution Replit built Cloud Agents that turn plain‑language requests into safe code changes, test them, and handle pull‑request reviews automatically.
Read on replit.comDetails -
ExpediaCoding agents
Beyond the Prompt: Engineering the Loop for Software Change at Scale
Problem Large code changes were slow because hidden dependencies and build failures required constant human intervention.
Solution Created a Loop system that combines prompts, reusable skills, and sub‑agents to automate iOS component migration, handling failures and decisions autonomously.
Two‑year effort finished in three months
Read on medium.comDetails -
AgodaCoding agents
Key Takeaways from Tech at Agoda: From Assistants to AI Agents
Problem Engineers wrote code faster, but planning, testing and release became bottlenecks.
Solution Agoda created AI agents that take on planning, code review, support triage and testing, using clear goals and human approval steps to keep work safe.
Read on medium.comDetails -
SentryCoding agents
How a beagle-shaped AI agent ships code to madewithvuejs.com
Problem Developers needed an easy way to generate and deploy code without manual steps.
Solution Sentry created a beagle‑shaped AI coding agent that writes, tests, and ships code through a unified CLI and sandboxed MCP environment.
Read on blog.sentry.ioDetails -
ShopifyCoding agents
We gave AI agents a mobile dev environment
Problem AI agents were overloading developers' laptops when building and testing mobile apps.
Solution Created Remote Dev, a cloud‑based Mac environment that gives each AI agent its own sandbox to compile, run simulators, and create PRs, all triggered from Slack.
Read on shopify.engineeringDetails -
NVIDIACoding agents
Running Frontier-Class AI Models Across 8 NVIDIA DGX Spark Nodes
Problem Developers need to run very large models and multiple models together, but single machines lack enough memory and speed.
Solution NVIDIA built an eight‑node DGX Spark cluster that links eight servers, giving 1 TB unified memory and faster communication, letting huge models run and be served locally.
80 tokens per second generation
Read on developer.nvidia.comDetails -
NVIDIACoding agents
Developing an AI Factory Digital Twin with an Agent-Assisted Workflow
Problem Developers struggled to combine site plans, equipment data, and engineering specs into a single, editable digital twin model.
Solution Created an Omniverse DSX Blueprint with reusable assets and scripts, and added AI agents that gather inputs, assemble OpenUSD scenes, and let engineers review and iterate designs.
Read on developer.nvidia.comDetails -
CloudflareCoding agents
How Commerce.com and Cloudflare are getting storefronts ready for AI agents
Problem Merchants needed a way for AI agents to shop on their sites without breaking when pages change.
Solution They added an edge script that injects structured tools into pages, letting agents call exact functions like search_products instead of guessing clicks.
Read on blog.cloudflare.comDetails -
WixCoding agentsComponents Are Dead; And Now for Something Completely Different
Problem AI agents find multi‑file UI components hard to understand and modify.
Solution Wix built a new frontend framework that flattens pages into local HTML regions while keeping logic in headless components, letting agents work on a single file and verify changes at build time.
Read on wix.engineeringDetails -
NVIDIACoding agents
Scale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72
Problem Developers needed an AI that could handle huge codebases and multi-step refactoring without breaking tasks into pieces.
Solution Created Muse Spark Open, an 816B MoE model with a 1 M‑token window that plans, uses tools, observes results, and self‑corrects across entire projects.
Read on developer.nvidia.comDetails -
NVIDIACoding agents
How to Rewrite Apache Spark UDFs for GPU Acceleration with AI Agents
Problem Custom Spark UDFs slowed down queries and prevented GPU acceleration.
Solution They created AI agents that automatically rewrite UDFs into GPU‑compatible code, letting Spark run fully on GPUs and boosting speed over 20×.
22x+ speedups
Read on developer.nvidia.comDetails -
WixCoding agents
Our AI Agent’s Design Looked Complete. 3 Runtime Checks Proved It Wrong
Problem AI‑generated web designs seemed finished but often broke when run in production.
Solution Wix added strict design‑log reviews, runtime checks, and self‑critiquing steps for the AI coding agent, catching over‑engineering and false assumptions before shipping.
Read on wix.engineeringDetails -
HelloFreshCoding agents
From Coding Assistants to Agentic Teams at HelloFresh
Problem Engineers struggled with slow pull‑request throughput and high manual coding effort.
Solution HelloFresh created AI‑driven coding agents that write, review and merge code, integrating them into product teams to speed up development and increase PR output.
Pull‑request throughput nearly doubled
Read on engineering.hellofresh.comDetails -
NVIDIACoding agents
Validate AI Factory Changes with Digital Twins and AI Agents
Problem Teams had to wait for physical hardware before they could test infrastructure and software changes, slowing deployments.
Solution They created a digital‑twin simulation of the AI factory and AI agents that query the twin, run checks, and trigger workflows, letting CI/CD validate changes early.
Read on developer.nvidia.comDetails -
GitHubCoding agents
Building Git infrastructure for agent-scale development
Problem Developers needed Git operations that could handle thousands of AI agents working on code simultaneously.
Solution Created a new storage layer and core services that scale to agent‑level traffic, using sharding, caching, and async pipelines to keep repository actions fast.
Read on github.blogDetails -
NVIDIACoding agents
Build Applications on NVIDIA BlueField Faster with NVIDIA DOCA Agent Skills
Problem General AI agents lacked specific knowledge of NVIDIA DOCA, causing guesswork and slow development cycles.
Solution Created DOCA Agent Skills that give agents verified API signatures, hardware requirements, and build constraints, letting them reason like expert DOCA developers.
Read on developer.nvidia.comDetails -
NVIDIACoding agents
AI Native by Design: Lessons Learned from Building NVIDIA TensorRT Model Connect
Problem Model developers struggled to get high‑performance TensorRT inference without deep expertise.
Solution Created an open‑source C++ library where AI agents generate candidate implementations, then isolate, test, benchmark, and review each change before shipping.
Read on developer.nvidia.comDetails -
Hugging FaceCoding agents
Holo4: powering generalist computer-use agents
Problem Businesses needed a single AI agent that could work with screens, code, and APIs without switching models.
Solution Holo4 was trained with supervised and reinforcement learning on many tasks, letting it click, type, write code, and call tools, and it runs on any platform with one model.
Much lower cost per task
Read on huggingface.coDetails -
NVIDIACoding agents
Accelerating a ROS 2 Node with an AI Agent and NVIDIA Isaac ROS
Problem Robotics nodes were copying data through CPU memory, slowing down GPU‑accelerated perception.
Solution An AI coding agent inspects a CUDA‑accelerated ROS 2 node, plans a minimal refactor to use rosidl::Buffer, and verifies zero‑copy GPU transport.
Read on developer.nvidia.comDetails -
NVIDIACoding agents
Translating CUDA Tile Operations from Python to Rust Using Agentic AI
Problem Developers needed Rust versions of many existing CUDA Tile kernels without rewriting them manually.
Solution Created an AI‑driven multi‑agent pipeline that reads Python or Triton kernels, translates them to Rust, validates each step, and benchmarks performance.
99.5% of Python performance achieved
Read on developer.nvidia.comDetails -
DoctolibCoding agentsKeeping AI on Track: Lessons from Doctolib’s Oxygen Design System
Problem Designers needed to create UI code quickly without engineers, but couldn’t trust AI‑generated code to follow the design system.
Solution Doctolib built the Claude Oxygen Plugin that feeds design‑system context to an LLM, enforces compliance rules, and runs evaluations, turning prompts into React code that matches their component library.
Read on medium.comDetails -
SourcegraphCoding agents
Introducing Agentic Batch Changes: the frontier agent for code change at scale
Problem Large codebases need coordinated updates across thousands of repositories, which is hard and error‑prone.
Solution Sourcegraph created an agent that takes a plain‑language change request, plans the edit, tests it in one repo, then automatically rolls it out and adapts to each repository, tracking PRs until merge.
Read on sourcegraph.comDetails -
ShopifyCoding agents
Migrating Shop app from React Native to native
Problem Maintaining a shared React Native codebase limited performance and added complexity for iOS and Android.
Solution The team used AI coding agents to translate React Native screens into native SwiftUI and Jetpack Compose, rebuilding the Shop app in 12 weeks.
Native app released in 12 weeks
Read on shopify.engineeringDetails -
NVIDIACoding agents
Frontier Reasoning Reaches the Edge: How to Deploy and Optimize Models on NVIDIA Jetson
Problem Developers could not run multi‑step reasoning agents on edge devices, forcing costly cloud calls and risking data exposure.
Solution They explain how to deploy compact open models like Nemotron 3.5 Lightning and Qwen3.8‑27B on Jetson, using distillation, mixture‑of‑experts and inference optimizations to fit memory and speed constraints.
Read on developer.nvidia.comDetails -
Hugging FaceCoding agents
Give Your Coding Agents a Memory You Own
Problem Coding agents forget previous work when a session ends, forcing users to re‑enter context.
Solution Created a local memory layer called funes that indexes agent session logs, chunks them, embeds them, and reranks results so agents can retrieve exact past turns during new tasks.
Read on huggingface.coDetails -
SourcegraphCoding agents
A smarter way to run code migrations with less LLM context
Problem Scanning thousands of files for migrations flooded LLM context, driving up token costs.
Solution Sourcegraph built Deep Search, a sandbox that runs scripts on search results, aggregates data, and returns concise CSV/JSON reports, feeding only the summary to the LLM.
Read on sourcegraph.comDetails -
ExpediaCoding agents
The Context Tax of Agentic Development
Problem Engineers and agents struggled because environment URLs hid the real backend, causing wrong work and delays.
Solution They added explicit, structured context that maps UI URLs to actual backends, so both humans and agents can verify and act on the correct environment.
Read on medium.comDetails -
OLXCoding agents
Does it matter how you feed an agent its context?
Problem Engineers ran out of token budget while using a coding agent for a complex backend task.
Solution They compared four ways to give the agent context—single prompt, handoff notes, etc.—and measured cost and performance for the same coding job.
Read on tech.olx.comDetails -
FaireCoding agents
Faire’s agentic development culture
Problem Engineers wanted to use advanced AI tools early without waiting for vendors.
Solution They built lightweight layers around existing models, encouraging voluntary use and showing benefits to drive natural adoption.
Read on craft.faire.comDetails -
RampCoding agentsIntegrations That Write Themselves
Problem Customers kept asking for new integrations, slowing down onboarding and product adoption.
Solution Created an autonomous agent that uses LLMs to write, test, and maintain integration code whenever a request arrives.
Read on builders.ramp.comDetails -
ZalandoCoding agents
Agentic Engineering at Zalando: a snapshot
Problem Engineers struggled to experiment with LLMs across providers and track usage costs.
Solution Created a LiteLLM API proxy with pre/post hooks for cost tracking, model retirement, and prompt‑caching injection, plus a chat UI and CLI, running on tiny pods.
Supports 2k MAU with six small pods
Read on engineering.zalando.comDetails -
FaireCoding agents
Spec, review, ship: How AI is changing the way Faire builds software
Problem Engineers struggled to turn vague ideas into working code quickly across many tech stacks.
Solution They built a shared specification workflow and AI agents that generate, review, and orchestrate code tasks, letting engineers focus on judgment.
600 pull requests merged
Read on craft.faire.comDetails -
AllegroCoding agents
How AI Agents Halved the Effort of Migrating 500 Gradle Modules from Koin to Hilt
Problem Migrating 500 Gradle modules from Koin to Hilt was slow and risked app performance.
Solution Allegro created AI agents that read code, suggest changes, and apply them, cutting migration time per module from a week to under two days.
Effort halved, migration time reduced
Read on blog.allegro.techDetails -
Hugging FaceCoding agents
From the Hugging Face Hub to robot hardware with Strands Agents and LeRobot
Problem Robotics teams had to juggle five separate tools to record demos, train, test, and deploy policies, slowing development.
Solution They created a single Strands agent that wraps LeRobot functions as tools, letting users record, push, simulate, and run policies on real robots with one codebase.
Read on huggingface.coDetails -
SlackCoding agents
Agentic Testing: Where Agents Fit in the E2E Testing Stack
Problem Testing teams struggled to cover all user paths in end‑to‑end tests.
Solution They built LLM‑powered agents that automatically generate and run exploratory E2E tests, then compare results with deterministic test suites.
Read on slack.engineeringDetails -
Hugging FaceCoding agents
How an Agent Built a 3D Paris Gallery by Chaining Two Hugging Face Spaces
Problem Creating a 3D Paris gallery required manually using image and 3D tools, which was complex and time‑consuming.
Solution An agent was programmed to call two Hugging Face Spaces: one to generate images, another to turn them into 3D Gaussian splats, then automatically combined the results into a web viewer.
Live static Space showcasing the gallery
Read on huggingface.coDetails -
ReplitCoding agents
Build a custom Shopify storefront on Replit
Problem Users wanted a fast, conversational way to create and launch a Shopify storefront without manual setup.
Solution Replit built an Agent that chats with users, provisions a Shopify store, generates a custom front‑end, and links everything, all from a single conversation.
Store created in about ten minutes
Read on replit.comDetails -
Hugging FaceCoding agents
Designing the hf CLI as an agent-optimized way to work with the Hub
Problem Coding agents had to use many tokens when calling the Hub via CLI or SDK, slowing down multi‑step tasks.
Solution They added automatic agent detection, created a special agent‑mode output without colors or truncation, and tagged requests, letting agents work faster and with less token usage.
Up to 6× fewer tokens used
Read on huggingface.coDetails -
ReplitCoding agents
Meet Replit SEO Agent
Problem Published apps were hard to find on web and AI search, limiting traffic.
Solution Created an SEO Agent that scans a live app, reports SEO problems, and applies fixes automatically, while adding semantic HTML and meta tags by default.
Read on replit.comDetails -
ShopifyCoding agents
Under the River
Problem Engineers faced slow feedback and difficult code changes across many repositories.
Solution They created River, a Slack AI agent, and rebuilt their infrastructure into a single repo with fast CI, caching, and reproducible environments so the agent could help with code tasks.
One in eight merged PRs co‑authored by River
Read on shopify.engineeringDetails -
ShopifyCoding agents
Flow generation through natural language: An agentic modeling approach
Problem Store owners find it hard to create Flow automations without coding.
Solution Shopify fine‑tuned Qwen3‑32B into a tool‑calling agent that converts plain English into Flow workflows, using merchant data and weekly retraining for speed and cost gains.
2.2x faster, 68% cheaper
Read on shopify.engineeringDetails -
ReplitCoding agents
How product managers ship faster using Replit's agentic workflows
Problem Product managers spent too much time manually syncing docs, decks, and tickets, leading to stale information.
Solution They created an agentic workflow that treats the prototype as the source of truth, automatically updating all related artifacts and keeping everyone aligned.
Read on replit.comDetails -
GoogleCoding agents
Vibe Coding XR: Accelerating AI + XR prototyping with XR Blocks and Gemini
Problem Creating XR experiences required complex code and hardware setup, slowing down experimentation.
Solution They built a workflow that lets Gemini turn a text prompt into a physics‑aware WebXR app in under a minute, using XR Blocks and a simulated desktop preview.
Prototypes generated in under 60 seconds
Read on research.googleDetails -
ReplitCoding agents
Live from Replit HQ Part 2
Problem Multiple AI agents editing the same code caused merge conflicts and coordination problems.
Solution They created Agent 4, a pipeline of specialized models that plans task order, runs parallel jobs, and uses coding models to automatically resolve most merge conflicts, surfacing only edge cases to the user.
90% of conflicts resolve automatically
Read on replit.comDetails -
ReplitCoding agents
What’s changed from Replit Agent 3 to Agent 4
Problem Developers had to choose an artifact type early and use a separate design mode, slowing and complicating app creation.
Solution Agent 4 lets users describe any kind of artifact, builds it with integrated tools, and offers a unified Design Canvas that shows live previews and mockups side‑by‑side, working for all artifact types.
Read on replit.comDetails -
MediumCoding agentsMaking AI Write Android Code Our Way: A Practical Guide to Agent Skills
Problem AI code suggestions ignored our Android team's specific conventions and patterns.
Solution We created AGENTS.md and custom Agent skills that teach the AI our project’s architecture, libraries, and step‑by‑step workflows, so generated code follows our standards.
Read on medium.engineeringDetails -
ZendeskCoding agentsWhat happens when code becomes abundant
Problem Developers were writing a lot of repetitive code, slowing product development.
Solution Zendesk created an AI coding assistant that writes, reviews, and tests code, letting engineers focus on higher‑level work.
Read on zendesk.engineeringDetails -
ReplitCoding agents
Introducing Replit Agent 4: Built for Creativity
Problem Developers spent too much time coordinating design and code, slowing creative work.
Solution Replit created Agent 4, an autonomous assistant that takes simple prompts, plans tasks, calls tools, and iterates code and design in the same environment, letting users focus on creativity.
Production‑ready apps shipped 10× faster
Read on replit.comDetails -
HubSpotCoding agentsAutomated Code Review: The 6-Month Evolution
Problem Code reviews were a major bottleneck, slowing engineers down.
Solution HubSpot built an AI‑powered code reviewer that runs on every pull request, using a Java‑based agent framework to read code and post feedback automatically.
Feedback time cut 90% (up to 99.76% in September)
Read on product.hubspot.comDetails -
ReplitCoding agents
Building Spookseek AR on Replit: How a Designer Shipped an AR Ghost Hunting Game in a Week
Problem A designer with no coding experience needed to create a real AR ghost‑hunting game fast.
Solution She used Replit Agent to write and run the 3D AR code, iterating daily and deploying a TestFlight build within a week.
Read on replit.comDetails -
HubSpotCoding agents
Cloud Coding Agents at HubSpot
Problem Engineers needed faster, reliable code generation and review without manual effort.
Solution HubSpot built cloud coding agents that integrate with GitHub, run tests, and use internal tools to generate and review code, handling thousands of pull requests.
7,000 AI‑generated PRs merged, 50,000 PRs reviewed
Read on product.hubspot.comDetails
No write-ups match these filters.