.svg)
From the podcast "Why Isn't This AI Thing Working?" — a conversation with Jim and Alysha Rosser, co-founders of Entry Studios and the team behind Grove.
Everyone was promised that AI would hand back their time. So why does it still feel like a full-time job to trust your AI agents — babysitting four terminals, clicking "yes, allow this" one prompt at a time, and re-reading long specs just to confirm the model did what you asked? On this episode of Why Isn't This AI Thing Working, Jess Lampe sits down with Jim and Alysha Rosser, co-founders of Entry Studios and the team behind Grove, to unpack why AI stalls the moment the pilot ends and real work begins.
Their diagnosis is refreshingly blunt: it isn't a capability problem. It's a trust problem. And until you fix that, all the raw speed in the world just moves your headaches somewhere else.
Ask most teams what's frustrating about AI and you'll get a pile of small complaints. Jim's insight is that nearly all of them fit under one umbrella: a lack of trust. Agents skip instructions you gave them, then cheerfully admit it after the fact. They invent a new way to run a command you've already defined. They stop halfway to pat themselves on the back for work that isn't finished. And when they do produce something, you're often left with more questions than answers — did this actually pass the tests, will it clear the CI/CD gates, is this even what I asked for?
That's the core of the problem. If you can't verify what an agent did, you can't walk away from it. And if you can't walk away, you haven't actually delegated anything.
Jim puts it plainly: "I'm not a micromanage person with humans — I certainly don't want to be with computers." The whole promise of AI is that it does the work while you go do something else. But the day-to-day reality is the opposite. You send an agent off with good intentions, run to grab lunch, and come back to find it sitting there, proud of a half-finished job, waiting on you.
Jess offers the perfect metaphor for that state: it feels like waiting tables. Four terminals open, bouncing between them, never off the floor. That's not delegation — that's just a different kind of busywork.
The difference between babysitting and delegating comes down to a single question: can you verify the work without watching it happen? Get that right and you can queue up sessions, close the laptop, and review real results in the morning. Get it wrong and you've simply doubled your workload.
Grove's answer isn't "use less AI." Both Jim and Alysha are emphatic that AI has genuinely accelerated their work. Their answer is to wrap AI in something the industry spent years perfecting and then quietly forgot: determinism.
Alysha frames it as a mindset she's carried since her services days — build compound AI systems that use the model only for the parts that need to be probabilistic, and wrap repeatable, deterministic process around everything else. Or, as she puts it, "make determinism cool again."
In practice that means task lists with hard proof requirements. You define each task you want done and the evidence you expect back: Did it follow test-driven development? Did every happy path and edge case get covered? Show me the linter output. Show me the passing test suite. Crucially, Grove — not the AI — runs those checks. It verifies the exit codes itself, because trusting the model to grade its own homework defeats the point.
The other half of trust is control. Everyone worries about the agent that deletes a production database, so most tools bolt on allow/deny permissions. Grove treats permissions as boundaries — and the distinction matters.
A plain "deny" just tells the agent what it can't do, so it burns time hunting for another way to do it. Grove's boundaries carry context: not just no, but use this Git MCP server instead, and here's why. Jess describes it as the difference between dumb bowling bumpers and intelligent ones — instead of the ball just clanging off the rails, it gets nudged toward the center pin. The agent self-corrects and gets to the right answer faster.
When you review AI output at scale, you need receipts, not vibes. Every Grove session produces an executive summary by default, and it mirrors your task list point for point. Want a before-and-after screenshot of that new export button? A short screen recording of the feature? An accessibility and brand audit? Annotated, human-readable explanations of the code that changed? It's all captured, so a reviewer can skim the evidence and approve with confidence that the change will clear the rest of the pipeline.
The proof of the philosophy is in how the Rossers actually work. They use Grove to build Grove — "a car that's building itself while you drive." They queue up sessions at the end of the day, sometimes with a designated "night manager" session to handle anything that would otherwise block progress, and walk in the next morning to a stack of finished pull requests.
The numbers are striking for a two-person team: roughly ten reviewable PRs a day, and around 900 merged since they started tracking in July. That's a pace most teams only aspire to — and it's possible precisely because the review step is fast and trustworthy, not because anyone is typing faster.
Underneath all of it is a theme Alysha returns to again and again: cognitive load. AI didn't eliminate it — it made it easier than ever to generate more of it. Fire off a vague prompt, get back a twenty-page spec padded with things you never asked for, skim it, pass it downstream, and now it's someone else's burden to decode. "We've become the least efficient communicators in the history of communications," Jim jokes.
The alternative is to spend that load up front, by design. Grove's Studio lets teams model an application the way they'd actually reason about it — data sources, models, logic, interfaces, personas — and link a plain-language view for product folks to a technical view for engineers, with a query language that flags structural gaps before a line of code is written. The vision Alysha describes is one where you spend most of your time designing, your schematics become the spec, and the conversation window is the exception, not the job. As she says: "If you're spending all your time in the conversation tab, you're under-using Grove."
That's the shift this episode is really about — moving from talking at AI all day to designing with intent, and finally being able to trust your AI agents enough to walk away.
Listen to the full conversation with Jim and Alysha Rosser on Why Isn't This AI Thing Working. Learn more about Grove at entrystudios.com.
From the podcast "Why Isn't This AI Thing Working?" — a conversation with Jim and Alysha Rosser, co-founders of Entry Studios and the team behind Grove.
Everyone was promised that AI would hand back their time. So why does it still feel like a full-time job to trust your AI agents — babysitting four terminals, clicking "yes, allow this" one prompt at a time, and re-reading long specs just to confirm the model did what you asked? On this episode of Why Isn't This AI Thing Working, Jess Lampe sits down with Jim and Alysha Rosser, co-founders of Entry Studios and the team behind Grove, to unpack why AI stalls the moment the pilot ends and real work begins.
Their diagnosis is refreshingly blunt: it isn't a capability problem. It's a trust problem. And until you fix that, all the raw speed in the world just moves your headaches somewhere else.
Ask most teams what's frustrating about AI and you'll get a pile of small complaints. Jim's insight is that nearly all of them fit under one umbrella: a lack of trust. Agents skip instructions you gave them, then cheerfully admit it after the fact. They invent a new way to run a command you've already defined. They stop halfway to pat themselves on the back for work that isn't finished. And when they do produce something, you're often left with more questions than answers — did this actually pass the tests, will it clear the CI/CD gates, is this even what I asked for?
That's the core of the problem. If you can't verify what an agent did, you can't walk away from it. And if you can't walk away, you haven't actually delegated anything.
Jim puts it plainly: "I'm not a micromanage person with humans — I certainly don't want to be with computers." The whole promise of AI is that it does the work while you go do something else. But the day-to-day reality is the opposite. You send an agent off with good intentions, run to grab lunch, and come back to find it sitting there, proud of a half-finished job, waiting on you.
Jess offers the perfect metaphor for that state: it feels like waiting tables. Four terminals open, bouncing between them, never off the floor. That's not delegation — that's just a different kind of busywork.
The difference between babysitting and delegating comes down to a single question: can you verify the work without watching it happen? Get that right and you can queue up sessions, close the laptop, and review real results in the morning. Get it wrong and you've simply doubled your workload.
Grove's answer isn't "use less AI." Both Jim and Alysha are emphatic that AI has genuinely accelerated their work. Their answer is to wrap AI in something the industry spent years perfecting and then quietly forgot: determinism.
Alysha frames it as a mindset she's carried since her services days — build compound AI systems that use the model only for the parts that need to be probabilistic, and wrap repeatable, deterministic process around everything else. Or, as she puts it, "make determinism cool again."
In practice that means task lists with hard proof requirements. You define each task you want done and the evidence you expect back: Did it follow test-driven development? Did every happy path and edge case get covered? Show me the linter output. Show me the passing test suite. Crucially, Grove — not the AI — runs those checks. It verifies the exit codes itself, because trusting the model to grade its own homework defeats the point.
The other half of trust is control. Everyone worries about the agent that deletes a production database, so most tools bolt on allow/deny permissions. Grove treats permissions as boundaries — and the distinction matters.
A plain "deny" just tells the agent what it can't do, so it burns time hunting for another way to do it. Grove's boundaries carry context: not just no, but use this Git MCP server instead, and here's why. Jess describes it as the difference between dumb bowling bumpers and intelligent ones — instead of the ball just clanging off the rails, it gets nudged toward the center pin. The agent self-corrects and gets to the right answer faster.
When you review AI output at scale, you need receipts, not vibes. Every Grove session produces an executive summary by default, and it mirrors your task list point for point. Want a before-and-after screenshot of that new export button? A short screen recording of the feature? An accessibility and brand audit? Annotated, human-readable explanations of the code that changed? It's all captured, so a reviewer can skim the evidence and approve with confidence that the change will clear the rest of the pipeline.
The proof of the philosophy is in how the Rossers actually work. They use Grove to build Grove — "a car that's building itself while you drive." They queue up sessions at the end of the day, sometimes with a designated "night manager" session to handle anything that would otherwise block progress, and walk in the next morning to a stack of finished pull requests.
The numbers are striking for a two-person team: roughly ten reviewable PRs a day, and around 900 merged since they started tracking in July. That's a pace most teams only aspire to — and it's possible precisely because the review step is fast and trustworthy, not because anyone is typing faster.
Underneath all of it is a theme Alysha returns to again and again: cognitive load. AI didn't eliminate it — it made it easier than ever to generate more of it. Fire off a vague prompt, get back a twenty-page spec padded with things you never asked for, skim it, pass it downstream, and now it's someone else's burden to decode. "We've become the least efficient communicators in the history of communications," Jim jokes.
The alternative is to spend that load up front, by design. Grove's Studio lets teams model an application the way they'd actually reason about it — data sources, models, logic, interfaces, personas — and link a plain-language view for product folks to a technical view for engineers, with a query language that flags structural gaps before a line of code is written. The vision Alysha describes is one where you spend most of your time designing, your schematics become the spec, and the conversation window is the exception, not the job. As she says: "If you're spending all your time in the conversation tab, you're under-using Grove."
That's the shift this episode is really about — moving from talking at AI all day to designing with intent, and finally being able to trust your AI agents enough to walk away.
Listen to the full conversation with Jim and Alysha Rosser on Why Isn't This AI Thing Working. Learn more about Grove at entrystudios.com.
