Mixed Mobbing: A new PAIR-adigm?

Pairing with Bots is about trying to apply some of the early agile practices to our new AI-assisted coding reality. I’ve been on this journey for almost a year now, and I have to tell you, it’s been an adjustment for me on a personal, and dare I say, psychological level. Among the most pressing questions weighing on my mind is:

How do we collaborate as humans when the focus is on AI?

While I haven’t completely figured this out yet… I think I can safely describe what’s working and what’s not working at this point. Among the things in the “not working” category are:

• 8–10 hour days intensely pairing with AI (not healthy)
• Asking co-workers to code review PR’s with 10+k lines of code
• Delivering to deadlines based on AI prototype code
• Working within existing design, planning and business processes
• Mentally “owning” AI-generated code as “developed by me”

These things have led to some real exhaustion, exasperation, emotional outbursts and profound disappointments. Fortunately, I’ve also been trying a lot of other things, and some of them seem to be working quite well, although it does require the cooperation of others, so gentle persuasion may be required:

• Working collaboratively with people for about half the day
• Real-time code reviews (an engaged human pair or group)
• A shared understanding that AI is driving, but we are navigating
• A willingness to experiment and keep what works (it’s cheap now!)
• Not tying my self-worth as a “programmer” to the code produced

For the remainder of this article, I’d like to talk about a very specific practice to enable some of these better approaches to play out, and hopefully change the way we work with AI assistance for the whole team, not just as an individual developer.

Most development teams that pair program have attempted “mob programming” once or twice… Some do it often. As the name implies, it’s bringing more than the usual 2 people together looking at the same screen in a cacophony of co-creation that can serve many purposes:

Being present “when the magic is created”

Software development is weird. Most people see a user interface or final report output, but the “magic” behind what’s actually happening is really, really hard to figure out and understand. LLMs are perhaps the pinnacle of this obfuscation. Very few truly understand what’s going on behind the scenes in order to extract human-like responses from electron-charged sand. Mobbing is the chance for less technical stakeholders to peek behind the curtain and see what the wizard is up to. Even if they don’t understand the source code… just overhearing the design decisions being made and the underlying technologies, interactions and components being used will almost certainly pay off later on.

Hedging against Employee Churn

“Bob’s the one who knows about this stuff”. Didn’t he quit last week? It’s good for more than one person to know how things are built at a company. It’s even better for more than one developer to be able to use and extend the software that you have without a reverse-engineering research project with a steep, long learning curve. You can instead amortize this learning over time by pair programming and mob programming. That way, everyone understands the “magic”, and new heroes can arise to avenge the fallen. (Just kidding, I hope your company doesn’t have a problematic “hero worship” ethos)

Mastering your Programming Language’s Tools

One of the best things about pair (or mob) programming is that it shows you how to actually accomplish the act of programming instead of just telling you what to do. You can attend a lecture, or read a programming book/article… even watch a scripted YouTube screencast, but it’s not the same. We humans have this area on the side of our brains called the “mirror neurons” where we are hard-coded to mimic the behaviors of others with great ease… We can watch a dance routine and instantly come up with a hilarious facsimile of the complex moves for the socials. Watching is one thing, but also being able to stop, ask questions, retry something differently… well… that’s a whole new ball game. This is what Pairing with Bots has given us.

By watching the AI code, we can learn an “actual” programming language while interacting in our more comfortable human language… only many, many times faster than with a human pair programmer. Most LLM models are now so fast you have to actually turn “off” the auto-approve settings just to be able to keep up, depending on your level of comfort with what’s being done. I would highly recommend you stay engaged and learn from what’s being done by the AI (ask why!). You’ll know stuff.

But what about the rest of the mob that don’t know (and don’t really even want to learn) the underlying programming language? How should they engage? Yeah, the prompts… they are the “language” your average folk need to learn in 2025. It’s important. It’s empowering. It’s efficient. We programmers can’t do all the prompting, we need a whole bevvy of folks surrounding us that are good at all the other things that need to happen at a company. We need more folks that are also extremely comfortable using AI assistance in their diverse jobs. A lot of the customers my company serves use spreadsheets, for example… up until the current moment, there hasn’t been much agentic help for those who have mastered that particular tool, but that is changing. Everyone at your company needs to be using AI at a level where they can predict how much scrutiny, testing, and re-work will need to be applied to the output of their AI assistants. Working together in a mob can help.

Teaching and Learning

Employment in the 21st century can be so danged unfulfilling. But teaching others what you know and learning from them may be one of the few remaining benefits of going into the office or collaborating remotely (if you must). Developers can be closed off to learning non-technical “soft skills” or business-domain processes, or UI design principles. Some only want to know the details of memory management for that hot new programming language. These folks may not find mob programming with less-technical people in the room to be satisfying, and that is a shame… Because you know who can handle the details of memory management? AI. Do not hasten the realization of a (self-fulfilling) prophecy that it will replace you. What AI is very bad at (because it has no life…) is navigating the human experience and writing software that serves the needs of actual people.

The reverse is also true… The non technical folk need to understand what developers are up against. I recently had a setback where the LLM did exactly what I asked, it appeared to me and others to be working, but when it was all plugged in as a complete solution, it had created a data filter that technically worked from a certain (AI) perspective, but worked like no human would expect. We had to disable a feature in production because of it, so I guess you could say I was taught an important lesson about working with LLM coding assistants. It’s hard not to take this failure personally as an AI-assisted programmer. By making it public, however, awareness can begin to develop that we all need to be on the lookout for subtle hallucinations as we work.

But what does a mob look like in practice?

You’re going to want people who know things you don’t in the same “room” (physically, I hope… but at least virtually) and looking at the same screen (or projected image). I will behoove you to not call this a “meeting”. It is not. Maybe set expectations that this is “an experiment”, or a “brainstorming session”. (Personally, my favorite phrase is “co-creation with AI and the developers”.)

If you made it this far in the article, it’s probably safe to say you are a developer or have a passing interest in the craft, so it’s going to be incumbent upon you to switch your role be to the “driver”, and allow the rest of the crew you have assembled to be helpful “navigators” directing your efforts. (But secretly, you’re actually just the “navigator”, prompting your AI assistant.) Make their wildest dreams come true before their eyes.

By bringing people of diverse responsibilities and backgrounds together and exercising and developing the code with AI right in front of them, we’re reconciling the business and technical sides of the software we create in real time. This just wasn’t possible before AI. Nobody has the patience to watch even the fastest programmer fine-tune the UI or fix the compiler errors to get things working… but you can probably continue to carry on your conversation while the LLM works its magic for a few seconds between prompts. This especially works if your prototypes are EXTREMELY SIMPLE: We’re talking simple like CLI generating text output or maybe simple static HTML pages… deploying your full application and using the UI is a no-go, keep it simple.

I feel like we as software developers are missing the key benefit of all this “Vibe Coding” nonsense… we can make throw-away prototypes very quickly, but grinding out reliable production code takes longer… if we only spend time in the grind there is a huge “opportunity cost” for ideas that could make things better… but seem too hard to chase. What if that idea of yours changed the game, and only cost you an afternoon? Most world-altering software is just a harebrained experiment with good timing.

Just yesterday, for example, I spent most of the afternoon pair programming with my CTO, trying to prototype something completely different than we’ve ever tried before. It was extremely satisfying. We brought in a data scientist, client success person, and others for periods of collaboration to firm up assumptions and help us make decisions.

This is what works for me. I work for a very small company, and collaborating every day with a broad array of people from my company was already a requirement. I’ve worked for behemoths, and I know this can be very challenging at a huge company… but maybe that’s the point. Think small, demand simplicity, collaborate intensely, get stuff done.

Mixed Mobbing: A new PAIR-adigm? was originally published in Pairing with Bots on Medium, where people are continuing the conversation by highlighting and responding to this story.