Why do coding agents make solopreneurs more overwhelmed?
More power, more context, more problems
Disclaimer: This case study is a work of fiction. Any resemblance to actual people, companies or products is coincidental.
Rohan runs SpeedLaunch, a lifestyle business that brings in ~$1M/year by selling an opinionated SaaS launch kit subscription and newsletter for technical founders. He’s been doing this for five years, after a decade-long stint at Google working on Cloud products. Eventually, he grew tired of working for The Man, and decided to try his hand at self-authoring his own independence. He’s never wanted SpeedLaunch to “take over the world”, but early-stage founders have trusted his libraries to take care of hundreds of annoying infrastructure decisions on their behalf. They eagerly open up his newsletter every Monday to get his perspective on the industry. SpeedLaunch takes care of auth, billing, cloud setup, deployment scripts, and so on, so that founders can focus on zero-to-one as fast as possible.
Suddenly, Claude Code burst onto the scene, and the models have been getting better at an impressive rate. Coding agents now let founders do, by themselves, most of what SpeedLaunch used to do for them. Rohan’s been wrestling with the implications of this new reality for his business, and therefore his lifestyle.
Some customers have indeed churned because coding agents have made SpeedLaunch redundant for them. Others are still paying for it because they’re too inexperienced at doing all this themselves. They either get totally blocked on some random edge case and can’t debug the vibe-coded slop. Or they just want the peace of mind of having someone with Rohan’s experience, to unilaterally make a bunch of boring yet critical decisions for them. His business isn’t totally screwed yet, but things certainly aren’t looking good.
There’ve been lots of objective signs that his business isn’t as healthy as it was. His Stripe dashboard isn’t showing the natural growth he used to have. Each subscription cancellation email sends a wave of anxiety rippling through his body.
His doomscrolling on X hasn’t been doing him any favors either. Each time he gets some sort of bad news, he tells himself he’ll go onto X for ten minutes to do “market research”. However, ten minutes usually turns into forty. Watching competitors claim record-breaking ARR for an offering that’s obviously slop leads him to feel even worse. So he furiously bookmarks interesting threads that he never comes back to, saves screenshots, makes notes in a markdown file called “speedlaunch-pivots.md”, and tries to reply-guy to a few big accounts to make himself feel better. It doesn’t help that everything keeps moving so fast! Every day there’s some new model launch, new harness, or VC-backed company reporting breakthrough AI ARR. The FOMO is unreal. In the past, whenever he got anxious about SpeedLaunch’s viability, he’d tell himself that he could easily get back into FAANG. However, the recent “AI layoffs” have started poking holes in that story. Essentially, he’s been feeling increasingly anxious and stressed with each passing quarter since LLMs have become mainstream.
Rohan’s made his career by Just Doing Things. He’d built an impressive track record of promotions at Google, and he bootstrapped SpeedLaunch without VC money into a decent lifestyle while all his FAANG friends doubted him. Just Doing Things is how he copes whenever his anxiety is triggered. He gets inspired by a few different ideas from X, cracks open his MacBook and fires off a few Claude Code terminals. Honestly, he’s also been having fun using Claude Code. The last time he felt this way was when he first learned to program. It’s so viscerally satisfying to vomit an idea into a terminal, have an LLM say “Certainly!”, and watch it tirelessly bring it to life. Running parallel Claude Code terminals has been pure euphoria. When he’s hooked in, he’s the god of his own private universe. Each git branch represents a novel possibility for SpeedLaunch that he could have only dreamed of before. He’s integrated Claude into every single aspect of his workflow, and is now shipping features at a rate he literally couldn’t have fathomed before. He’s always on top of new automations, and has an agent constantly generating and reviewing code in the background, with him supervising it.
However, he’s started noticing cracks from this approach. His codebase has started to get kind of messy, hard to understand and hard to debug. Customers have started complaining that SpeedLaunch has been getting too bloated and confusing. They’ve often started privately telling him that his newsletters have lost their “voice”, perfectly coinciding with him using Opus to generate them.
They’ve started reporting bugs in parts of the codebase that absolutely need to remain stable and boring. At first, he chalked these problems up to his own bad prioritization and insufficient structure within his work. He tried formal sprint planning to get things under control. The experience immediately triggered a low-grade PTSD from his time at Google. BigTech really didn’t agree with him. All those pointless daily standups, inane planning spreadsheets and excessive overhead. Each formality that he introduced within SpeedLaunch triggered constriction and nausea in his body. Nevertheless, he tried to push all this down, but still kept getting distracted by random ideas from X.
All of this crashed into clarity due to a feedback call with a loyal customer named Priya. She’s a founder who’d used SpeedLaunch to bootstrap a few different projects. She’d believed in him when no one else did, and was one of his earliest customers. She’d trusted him enough to send him useful bug reports, helped him reproduce middle-of-the-night production issues and sent him lots of referrals.
“I’m sorry Rohan, but I need to cancel my subscription,” she said. “I signed up for SpeedLaunch because I trusted your taste and judgement. I don’t want to mess around with config files or deployment boilerplate when I’m trying to get something off the ground. But lately, SpeedLaunch has been so unwieldy. There are too many options, the docs are too confusing, there are AI features I don’t understand and bugs in places that need to be stable. I’ve felt cornered into slowing down and paying attention to infrastructure, when I’d like to be thinking about validating my business idea.”
He felt like he’d been punched in the gut, and was too dizzy to stand. For a few minutes, he could barely hear the rest of the call.
He used to make fun of the typical Silicon Valley tech bro founders who built solutions in search of problems, generated massive hype to raise VC, and built what would inevitably become enshittified products.
He’d promised himself that he’d never become that guy when he started SpeedLaunch.
Was his fixation with AI turning him into the worst version of himself? That’s certainly what Priya’s feedback sounded like, and it was consistent with what he saw on his Stripe dashboard, customer feedback tickets, etc.
How could his fears have blinded him so thoroughly, that he was on track to enshittify SpeedLaunch??
As the call ended, Rohan immediately recalled how he’d felt the first few weeks after Google. Founders came to SpeedLaunch, to him, because they trusted him. They trusted his taste, judgement and discernment. He’d quit his job as a cog in Google’s machine so he could build things customers genuinely loved and trusted. Things that would tangibly empower them to live the life of freedom he desperately wanted for himself too. He wanted a business that would let him and his customers make good money, but not sacrifice their sovereignty to the machine.
Rohan felt stuck between a rock and a hard place. He could keep rapidly exploring new ideas, or he could overpower his fears and just commit to something. The former felt extremely tempting because it was true that the market was changing, and would continue to change quickly. But it clearly wasn’t working. The feedback was clear that his rapid experimentation was making SpeedLaunch too bloated, incoherent and buggy.
However, forcibly committing to something that didn’t feel right also created stress. If he was “wrong”, he’d burn a lot of money, tokens and emotional energy in the process.
This internal conflict was tearing him apart, and he realized it’d eventually burn him out into a crisp.
Despite being a “doer”, he did something uncomfortable that challenged every instinct in his body. He slowed down to journal out his thoughts. He asked himself, “Which questions about SpeedLaunch am I avoiding that, if I engaged with them, would improve my life?”
SpeedLaunch had worked pre-LLM because customers could borrow his taste, judgement and discernment without needing to understand every decision behind it. It relieved them of needing to make a number of extremely mundane infrastructure decisions, to free them up for focusing on zero-to-one. It allowed him to continue investing in SpeedLaunch without sacrificing a sense of coherence between its features, customers and GTM. This coherence had grown consciously and unconsciously over the five years he’d been doing it.
Coding agents rapidly smashed his business’s coherence faster than Rohan could reconstitute it. Simultaneously, they opened up a substantial range of new possibilities, customer problems and strategic directions. Rohan drowned himself in this sea of possibilities he could execute on, but related to them via a purely technical lens rather than a broader strategic lens. For example, he kept asking himself what sorts of technical problems Claude Code could help him solve that he couldn’t solve before. Was prompt-to-SaaS actually viable? Or maybe prompt-to-intelligent-business-dashboard? Which of those would go viral? How could he optimize the number of parallel agents he spun up? Was he spending enough on tokens to truly maximize his productivity?
Essentially, Rohan had confused power with leadership. In my executive coach Brian Whetten‘s terminology, AI substantially increased Rohan’s power. It gave him the capacity to translate far more possibilities from potential ideas into concrete artifacts. All it took was his credit card and a new terminal tab with Claude Code. However, it didn’t eliminate the context and cognitive overhead that came with each one. Each git branch represented some implication for a customer archetype, an implicit/explicit promise of responsibility that SpeedLaunch would take on, an ongoing support burden, the potential introduction of new bugs, and an overall complexification of SpeedLaunch’s story for value creation.
This was especially painful when these branches pulled the product in competing directions. Often, Rohan would need to do further work to integrate all of this context together into a unified whole.
In Brian Whetten‘s terminology, leadership is the capacity to hold space for all of these competing contexts and possibilities, to create provisional clarity for the organization. This clarity then allows the organization to take responsibility for delivering specific future results with integrity. A leader needs to have the capacity to create structures like prioritization, goals, sprints, roadmaps, and so on, all in service of such clarity. A leader also needs to have the psychological capacity to sit in a sea of competing possibilities, that might seem like a set of polar opposites, without getting overwhelmed or anxious. They need to remain non-reactive despite any triggers that may arise from social media, customers, and so on, so that they can continue making wise choices with clarity.
Rohan’s power rapidly grew faster than his leadership capacity, in proportion to improvements in the underlying frontier models. As of this writing, o3 is a bit more than a year old. It feels like such an antiquated model now, but it blew Rohan’s mind when it first came out. Eventually, the competing branches of context he could rapidly generate with the latest models far outstripped what he could confidently hold, provisionally resolve, and translate into customer value.
According to Brian Whetten‘s terminology, Rohan’s stress was proportionate to the rapid increases in power that the models gave him, relative to his capacity for leadership. That is, Stress = Power - Leadership.
This stress was far from benign. It worsened his reactivity, decision-making and ability to hold context. It impaired his nutrition and sleep, which of course amplified these dynamics. He kept trying to solve this problem by maxing out token spend (i.e. increasing his power), because he didn’t understand this distinction between power and leadership. He kept spinning up more parallel branches, kept making more demos, and even got Claude to review all his PRs. This kept increasing his LLM bill, introduced bugs faster than he could resolve them, confused customers, and led to more churn, all of which contributed to his stress, and trapped him in a vicious feedback loop of stress.
In the background, the models kept getting substantially better every quarter, amplifying this vicious cycle of reactivity and stress. Each model upgrade silently and substantially increased this disparity between his power and leadership capacity.
Just because SpeedLaunch could do more, didn’t mean that his customers could understand it, trust it or receive value from it. He needed more leadership capacity to consciously choose which future SpeedLaunch would stand behind, so that it could successfully lead his customers into it.
He needed to create clarity for himself about what SpeedLaunch would stand for before he asked Claude Code to do anything else. He needed to bring SpeedLaunch’s purpose into explicit consciousness. AI was changing far too quickly for him to find a new purpose once, and then coast for another five years. Those days were gone. He needed a living practice to iterate on clarifying the value SpeedLaunch would take responsibility for delivering.
It took him a while to turn things around, but he eventually succeeded. The first six weeks were especially bumpy. Rebuilding trust with Priya wasn’t easy, but he got it done. Some prospects and randoms on X still asked him to build flashy but misaligned “AI-native” features. He politely ignored them. He started articulating an increasingly coherent and purpose-aligned story about who SpeedLaunch was, and how it’d solve its customers’ problems. He made it an active practice to say “no” to misalignments to keep his story coherent. In a sea of online slop, he started standing out and pulling in business rather than needing to push to get it. For the first time in a long time, he felt hope and felt joy at work.
He wasn’t out of the woods yet, but he could see a path.
The process that got him there was “simple” to describe, but “hard” to do. He ran it as an iterative loop that fluidly moved up/down through the steps, as he came into contact with disconfirming information.
He journaled until he could articulate a provisional purpose for SpeedLaunch that created a full-bodied “hell yes” for him. The old purpose had been something like “Help technical founders skip SaaS setup to launch faster”. He still cared about helping founders, but LLMs meant that selling artifacts themselves was no longer enough for a solopreneur like him. He tentatively up-leveled his purpose from selling artifacts to taking ownership of outcomes. “Help technical founders get from idea to customer validation via trusted and opinionated software infrastructure, even as code generation rapidly becomes free.”
He tested that provisional articulation of purpose in customer conversations. He’d already read The Mom Test and started talking to founders one at a time, exploring problems one at a time, to investigate what sort of value he could create for them. He paid attention for any patterns that emerged, and iterated on his stated purpose as necessary.
Once he’d convinced himself that there might be a “there” there, he translated his consciously articulated purpose into a number of specific goals. The specificity from these tractable goals allowed him to cleanly measure whether he was making progress toward his purpose or not. Specifically, it allowed him to measure whether his activities would create value for his customers or not, along his chosen purpose. He found that the ideal tractable goals were those that made sense to his intuition and reason, but felt a bit scary in his body.
Finalizing these tractable goals allowed him to create a roadmap of concrete tickets. Consciously creating coherence between his roadmap, tractable goals and broader purpose allowed him to more easily triage new ideas that came up. It made it a lot easier to resist the temptation to fire off some new experiments whenever he got triggered on X.
Rohan’s biggest obstacle to executing each step was his relationship with his emotions, and the walls they created between him and the life he wanted. This was his fundamental blocker to growing his capacity for leadership, and therefore metabolizing any new possibilities from AI.
He realized that if he wanted to become truly “AI-native”, he’d have to find a way to tether the growth of his leadership capacity to the underlying improvements in frontier model performance. But that’s a story for another day.
For now, he’s brought life back into his lifestyle business.
If any of this resonates with you, I’d love to hear from you at varun@doubleascent.com! I love talking to founders working through this.
Acknowledgements
Thank you to Brian Whetten for helping me internalize this distinction between power and leadership, and for helping me grow my capacity for leadership.