I recently asked a startup CEO for his take on the best AI coding tools. He mentioned the usual suspects: OpenAI’s Codex and Anthropic’s Claude Code.
Then he brought up a less familiar name.
“There’s another coding agent, Code Puppy, built by an amazing guy called Mike Pfaffenberger at Walmart, ” this CEO told me. “It’s got masses of usage in Walmart.”
The tip sent me down a rabbit hole that revealed something much bigger than another AI coding assistant. Code Puppy is part of Walmart’s effort to avoid what many technology executives fear could become one of the defining business problems of the AI era: getting locked into a handful of powerful providers.
The risk is familiar. Companies rush to adopt a breakthrough technology, redesign their systems around it, and then discover they’ve become dependent on a small number of suppliers. Switching becomes too expensive, disruptive, or risky.
It happened with IBM. It happened again with cloud computing. Now, many companies worry that it could happen with AI.
Code Puppy is Walmart’s attempt to avoid that fate. It’s become even more important lately, as companies blow through tech budgets by spending millions of dollars on AI coding tools and agents.
Pfaffenberger’s flex
Code Puppy was created by Pfaffenberger, a distinguished engineer in Walmart’s Global Tech group. The AI coding assistant helps developers write, edit, test, and manage software using natural-language instructions.
Like Claude Code and Codex, it can build features, fix bugs, and analyze projects. But unlike many rivals, Code Puppy isn’t tied to a single AI model or provider.
Instead, it can work with dozens of models from different suppliers, allowing developers to switch between them, compare results, or use several at once. It can also distribute workloads across providers, helping avoid usage limits and control costs.
That flexibility is central to Pfaffenberger’s vision.
“It gives us the ability to not be locked into a vendor and have freedom to control and integrate with our own internal systems,” he said during a presentation posted on YouTube in late April.
Cost control
This approach has multiple benefits. First up: Code Puppy could potentially save Walmart money on tokens, the main unit of AI usage. If one AI model provider increases token prices or introduces stricter rate limits, the system can help developers switch to cheaper models relatively easily.
Code Puppy works with models from OpenAI, Google, Anthropic, and dozens of other providers, according to the project’s public Github page.
And instead of sending every request to the same AI model, Code Puppy can be set up to automatically rotate between multiple models. That spreads workloads, reducing the risk of hitting rate limits.
Code control
Code Puppy is also about control, particularly over codebases, the vast collections of software code that underpin most modern companies.
AI coding tools like Claude Code and Codex are helping companies generate software at unprecedented speed, causing codebases to grow beyond what human developers can realistically maintain on their own.
That creates a potential dependency: if a codebase was largely built with Claude Code or Codex, companies may find they need to keep paying for those same tools to maintain, update, and understand the software they’ve created.
At first, Pfaffenberger said Code Puppy was a little more expensive than just using AI coding services such as Cursor or Windsurf.
“But what I really, really liked about it was I was in control,” he said during his recent presentation. “I was willing to pay a little bit more money to have my own source code base that nobody can mess with.”
“Enshittification”
At the moment, Anthropic and OpenAI are competing intensely with Cursor, Google, and others for AI coding market share. That means these services are still relatively affordable and flexible.
However, Pfaffenberger warned this could change, given how previous technology waves evolved.
In his presentation, he described this cycle of adoption and lock-in as the “enshittification” of tech. He showed a slide describing the process and how AI is intensifying it this time round.
Business Insider found the slide on CodePuppy’s public GitHub page. Because the presentation was stored as code instead of a PowerPoint deck, BI used ChatGPT to generate a viewable image of the slide.
Code Puppy public Github page
Enshittification, coined by writer Cory Doctorow, describes how technology platforms often become less attractive to users over time as companies seek to increase profits and exert more control.
“I’m really, I’m kind of proud of what I built here and just proud to be separated from what I would say is the investor-funded slop cycle,” Pfaffenberger said in his presentation. “That’s kind of like the gist of this slide.”
“Helpless”
Pfaffenberger’s concerns stem partly from his own experience as an enthusiastic user of AI coding tools. He said he began building Code Puppy after watching turbulence in the market for AI coding services.
Last year, when Windsurf appeared close to being acquired by OpenAI, Anthropic pulled access to one of its most popular models from the platform. Around the same time, Cursor sharply reduced usage limits, making heavy use of its service substantially more expensive.
“I was looking at all this happen and I felt kind of helpless,” Pfaffenberger said during the presentation.
So he built his own alternative. The first version took only a few hours to create. Pfaffenberger then used the software to improve itself, allowing the coding assistant to help develop new versions of Code Puppy.
Going viral inside Walmart
The project struck a chord inside Walmart. The Code Puppy team was given an award during a Walmart tech all-hands meeting.
And in a LinkedIn post, Walmart SVP Dave Glick said Code Puppy had spread well beyond engineering teams, with everyone from tech VPs to store managers using it to create simple automations and bring new ideas to life.
Pfaffenberger said Code Puppy “kind of went viral inside of Walmart.”
Qian Li, cofounder of startup DBOS, highlighted Pfaffenberger’s advocacy for using an “LLM council.” The idea is simple: instead of trusting one AI system, ask several models to tackle the same problem and compare their answers.
That approach reflects a broader philosophy behind the project. Rather than relying on one AI provider, Walmart can maintain flexibility to switch models as prices, performance, and capabilities change.
“Nightmare scenario”
Pfaffenberger is unusually outspoken about why that matters.
In the same presentation, he argued that the AI industry has become a circular ecosystem. AI model companies raise money to buy computing power from Nvidia. AI application startups raise money to buy access to those models. Users receive heavily subsidized services whose economics may not yet be sustainable.
Eventually, Pfaffenberger warned, somebody has to pay the bill.
“I just see this, like, nightmare scenario where… we as users who are not part of this agentic AI bubble, we need to do everything we can to protect ourselves from when that bubble goes pop and there’s no access to tokens or software or whatever,” he said in the presentation. “That was what motivated me to write Code Puppy.”
“Oh, I forgot, I forgot this part,” he added. “Walmart does require me to say that, like, all of my opinions are my own.”
“We build our tools to be platform agnostic, which gives us the flexibility to work with the best partners and capabilities across the industry. Our strategy is not to lock ourselves into one vendor or model, but to give associates access to the right tools for the right work as the technology continues to evolve,” a Walmart spokesperson said in a statement.
Sign up for BI’s Tech Memo newsletter here. Reach out to me via email at [email protected].
