Local Engineer Runs Out of Tokens, Discovers "Programming"

Robbie and Adam reopen the episode by tasting Redemption Rye, which is 46% alcohol, 95% rye, 5% barley, aged at least two years in new charred oak, and priced around $35. They smell black pepper, cola, corn, and strong alcohol, but find it sweeter, less spicy, and more approachable than Sagamore Rye. Adam rates it 6.25 tentacles; Robbie gives it 5.5. Robbie logs the rating in his personal assistant and says he wants the assistant to use accumulated preference data for whiskey, banjo, guitar, and birthday recommendations instead of merely storing information. By the end, the rye tastes harsher and less maple-sweet, so Adam lowers his rating to 5.5.

Robbie describes his workplace abruptly disabling all GitHub Copilot access so employees would use the internal AI product, Bob. The shutdown happened while people were actively working, and Robbie struggled to connect Bob to OpenCode after an authentication plugin failed. He tried taking over the plugin project and spent hours blocked by branch-protection rules: the main branch was protected, and any branch with “release” in its name was treated as immutable. Just as he finished untangling the problem, Copilot was restored until Friday. Robbie says a planned roll-off would have been less disruptive than suddenly pulling the tool, especially with only a few weeks left before the project shipped. Bob now performs better than its earlier beta version, but the timing remains poor.

Adam says an agent using browser screenshots, Chrome DevTools Protocol, Playwright, and a local Shopify Admin instance was supposed to copy styles and design tokens from one component to another during an overnight run. It claimed completion but stopped halfway, and Adam ultimately hand-wrote the CSS after another 45 minutes of arguing with it. Asked whether sub-agents would help, he says the task used one main agent looping to completion. Robbie observes that small prompts often work while a huge PRD and extensive design history overwhelm the model. They reject spectacular social-media claims that a single prompt reliably produces polished games or cinematic websites, arguing that online demonstrations often conceal substantial iteration and selection.

Adam suggests the failure may come from retaining too much context and proposes an orchestrator with specialized sub-agents. He describes a Shopify workflow he built around worktrees: a new worktree begins with a specification containing the branch’s goal and requirements, and the agent updates that document when requirements change. The spec becomes a durable, revisitable source of truth that allows Adam to clear the agent’s context and restart with focused information rather than preserving a long, confused conversation. Robbie agrees that large “dark factory” pull requests can be overwhelming, even when their 28 changed files are only agent instructions. They discuss routing files and skills as lightweight tables of contents so models discover only the information needed for the current task.

The hosts argue that the era of simply maximizing tokens is ending. Robbie cites reports that Amazon encouraged employees to use as much AI as possible and ranked them on a leaderboard, prompting people to generate useless code to appear productive. He also mentions a law firm offering $1 million for the highest token usage before finding no beneficial output. Shopify removed its leaderboard and retained only a usage chart, preserving visibility without rewarding a high number. Robbie still sees value in deliberately giving agents more work, but questions whether current token budgets are sustainable enough to replace teams. He predicts a cycle of cheap hosted models, expensive usage, local models, outdated hardware, and eventually engineers being hired to repair poor AI-generated systems.

Robbie and Adam debate whether models will become so capable that harnesses no longer matter. Adam notes that OpenCode currently gives him better results partly because its workflow feels more deliberate. Robbie is less optimistic: governments may restrict the strongest models, companies have incentives to limit supply and keep prices high for investors, and a truly strong open-source model could disrupt the market. They agree that Fable seemed to think harder and work better before its government-related interruption, but its return felt closer to Claude Opus. On code quality, Adam says “slop is a choice,” while both note that pre-AI code was already spaghetti. Robbie adds that old code becomes poor as browser standards and platform capabilities change, so AI is not uniquely responsible for mess.

Adam argues that frameworks may be over-engineering the web now that LLMs can write repetitive HTML, CSS, and JavaScript. He expects more server-rendered forms and thinner clients because many framework ceremonies solve problems that native platform features already handle. Robbie agrees that many sites update only an unread count occasionally, not millions of times per second, and says direct DOM operations could be sufficient for much of that work. Adam pushes back that frameworks historically reduce platform complexity and help teams scale, but their abstractions become liabilities when developers need the final 20 percent of flexibility. Their shared conclusion is that user experience matters more than preserving a particular code structure, and future tools should work closer to the platform.

Adam dismisses “meta-harnesses” that choose a harness, which chooses a model, calling them another layer of over-engineering. He says the important skill is articulating what should happen, not memorizing implementation details; he is trying to teach his children to write ideas down clearly. Robbie proposes a future where changing a specification immediately rebuilds the UI, restoring the feedback and motivation that coding currently provides. Adam says he is building Conjure around behavioral-driven design, spec-driven design, and service design: users describe flows and expectations, the tool generates journeys, requirements, design options, and eventually applications. It is intended to keep accessibility and internationalization in view from the start rather than bolting them on after narrow human decisions.

After discussing the Fourth of July, Robbie reports that Jeff, his GPS-guided robotic lawnmower, is failing badly. Its positioning can be four meters off, causing it to mow already-cut areas, wander into the woods, and drive off a rock cliff. When it loses its location, it repeatedly hits obstacles instead of backing up and recalibrating. Adam compares the experience with Roombas that leave their docks, eat socks, and accomplish nothing, saying he becomes unusually angry at malfunctioning inanimate objects. They joke that robot makers need to solve basic obstacle behavior before giving machines stronger motors, recalling demonstrations of robots lifting heavy weights and imagining a mower or household robot destroying something while insisting it is in a clear area.

Robbie says he and his wife bought AMP.ai, a workout machine positioned as a cheaper competitor to Tonal. It was eligible for his HSA and costs less than half as much as Tonal both initially and monthly. After installation, they completed benchmark workouts in which the machine progressively increased resistance to establish appropriate levels. The system can vary resistance throughout a movement, offer competitive modes, and learn workout preferences. The compact device mounts to wall studs and uses a phone on a dock as its screen, requiring power and an internet connection; Robbie was unsure whether the connection was Bluetooth or Wi-Fi. Adam summarizes it as a smart, electronically adjustable Bowflex, and Robbie says its first use went better than Jeff’s.