Nobody Knows What Their Job Is Anymore. Good.

Early in my career I worked at a founder-led software company. The founder understood the tax domain, knew the product inside out, could design the interface and write the code. All of it lived in one head.

We used to joke it was fragile. If he got hit by a bus, the whole thing went with him. CI/CD pipelines barely existed. Automation was primitive.

But here's what I remember most clearly: things got done fast. No handoffs, no translation loss, no reviewing work that had drifted from the original intent. Communication was instant because it was internal. One person, one mental model, zero gaps.

What we later called Agile (the standups, the sprints, the retros) was really just an attempt to recreate that speed at team scale. Communication infrastructure to compensate for knowledge being spread across multiple people.

Agile was never really about process. It was always about communication.

---

The Pipeline Has Always Been a Communication System

Before modern version control, teams used tools like Visual SourceSafe or SVN. File locking was real. If one developer had a file checked out, nobody else could touch it. Someone would shout across the office to ask if a module was free yet. Merging code was an event you scheduled and braced for. The tooling forced communication because the alternative was hours of untangling conflicts by hand.

Git made branches cheap and merges manageable. Real progress. But it also quietly removed a forcing function. When the friction of integrating code went away, some of the communication around it went away too. Teams got better at merging code and sometimes a little worse at merging thinking.

Then everything accelerated again. I saw a post recently from a developer running six AI coding agents in parallel, each in its own Git worktree, its own branch, its own pull request, all building different platform integrations in a single afternoon. Six agents. Six PRs. One afternoon.

Coding agents don't just accelerate the building step. They collapse the pipeline entirely. Someone with a clear idea can have a working prototype in an afternoon, without writing a spec, without waiting for a designer, without filing a ticket.

When you collapse the pipeline, you expose what was always underneath it. The bottleneck was never implementation. It was communication.

---

The Real Risk Nobody Is Talking About

A lot of teams are about to make a painful mistake.

When AI agents get added to an existing team, the assumption is that everything gets faster. More output, same people, same process. That's not what happens.

The bottleneck moves. Hard.

Before, implementation was the constraint. One engineer, one feature, takes however long it takes. Everyone paces themselves around that. Now anyone can generate a working prototype. A PM can build before the sprint starts. A designer can produce functional code directly.

Three or four times as much work to review, critique and decide on, handled by the same people at the same human pace.

Every role is quietly splitting in two. Writing test scripts, translating meetings into Jira tickets, producing mockups from a brief. That half is being absorbed by AI. Knowing which edge case breaks in production, spotting the gap between what a client asks for and what they actually need, understanding why a workflow feels wrong before anyone can articulate it. That half just got more valuable than ever.

The problem is most teams haven't figured out which half they were paying for.

The builders get faster and feel productive. The reviewers get buried and start to feel like the problem. The people generating the most output wonder why the team is still slow. The reviewers feel like gatekeepers blamed for a problem they didn't create.

Adding AI agents without rethinking how you communicate and review is like widening one lane of a motorway. The cars at the front just back up harder further down the road.

---

What AI Reveals About Communication

AI reveals the quality of your communication.

Vague prompts produce vague code. If you can't articulate clearly what you want built and why, the agent produces something. It just won't be the right thing. The cost of muddled thinking used to be absorbed by the time implementation took. Now muddled thinking ships fast.

Good product has always been good communication. The best PMs I've worked with were never the ones who wrote the most detailed specs. They had the clearest mental model and could transmit it without losing anything in translation.

The same was true of the best developers I worked with. We built software for accountants and SMSF trustees managing real money and real compliance obligations. Most developers kept that world at arm's length. Tax law, contribution caps, pension calculations. Not something you pick up on a Friday afternoon.

But some of the best ones did something that still stands out. They set up their own SMSFs. Not primarily as a tax strategy. They did it to feel the product from the inside. To sit in the same chair as the accountant at year-end, or the trustee trying to make sense of their own balance. They wanted to know where the software made people feel stupid.

It showed up in the questions they asked before writing a line. "Why would an accountant end up on this screen at year-end?" "What are they worried about getting wrong?" Those questions changed what they built. Over time the software felt different when people like that had touched it.

That founder I worked with early in my career had this quality in its purest form. He didn't need to communicate his mental model to anyone because he was the whole team. When I watch the best builders work with AI agents today, I see the same thing. A precise mental model, transmitted cleanly, with almost no loss.

AI doesn't change that truth. It amplifies it.

---

Everyone Thinks Their Role Is Most Advantaged. They're All Right.

A post went semi-viral recently describing the person most advantaged by coding agents as someone with an intuitive grasp of the product, genuinely bilingual between culture and technology, able to separate what's real from what's noise.

PMs shared it thinking it described great PMs. Designers shared it thinking it described great designers. Engineers shared it thinking it described great engineers. BAs shared it thinking it validated everything they'd been saying about requirements for years. Testers shared it and quietly wondered whether it was a compliment or a redundancy notice.

I shared it as a PM, fully confident it was about PMs. The designers in my network shared it the same day, fully confident it wasn't. We were all correct and none of us were listening to each other, which proves the point entirely.

They were all right. About the first part, at least.

Background matters less than it ever has. The archetype isn't a job title. It's the ability to hold a clear mental model and transmit it without loss. It can come from any discipline.

The one-person team was always the fastest unit in software. Not because of raw output, but because communication was instant. One person, working with agents, is starting to approach that speed again at a scale that actually matters. When every role splits into its thinking half and its mechanical half, one person holding the full mental model can let agents handle the mechanical side of all five roles at once. No translation. No handoffs. No loss.

Patrick and John Collison, Stripe's CEO and President, reported in their 2025 annual letter that the new startup cohort is growing 50% faster than the year before, with companies hitting $10 million revenue within three months of launch at double the prior year's rate. That data sits behind $1.9 trillion in real transactions. It's not a prediction. It's already happening.

The teams that figure out how to recreate that clarity across multiple people, shrinking the translation loss without shrinking the team, are the ones that will pull ahead.

That's the problem worth solving. It always has been.

---

Is your team faster since adding AI, or just busier? Most haven't figured out which one they are yet. That question shaped how we built BuildStability. Trainers already work hard enough. The AI should handle the admin, not create more of it.