Your AI Agents Don't Need You Anymore

Robbie and Adam taste Oak and Eden’s 100-proof bourbon with a five-inch Amburana spiral. The high-rye bourbon is made from 60% corn, 36% rye, and 4% malted barley, but the Brazilian wood dominates the flavor: cinnamon, Big Red, wassail, apple, coconut, and vanilla. Robbie says it resembles his whiskey-heavy mulled cider, while Adam describes a vanilla-sprinkled bear claw. They agree it is sweet and intense enough for perhaps one drink rather than six. Adam rates it 6.1, cautiously, and Robbie gives it 7.1; both want to try the custom builder.

Robbie describes Doorway, his nearly finished first app for people with ADHD or anyone who walks into a room and forgets why. A user can enter a multi-part task such as getting a hammer, screwdriver, and soda, and the app periodically asks whether everything was collected. A spouse can also add an instruction remotely, such as Caitlin telling Robbie to bring her water bottle, because he routinely returns empty-handed. The app is available for beta testing at doorway.help, which redirects to whydidiwalkinhere.com. Robbie says the problem feels worse in a world of phones, internet, and constant distraction.

Robbie says he has been using OpenCode, Claude’s free usage, and a desktop workflow to make agents do repeated work, but an orchestration loop quickly consumed his plan’s limit. He is productive while merely watching agents work, yet the speed creates a false impression that software is finished. His Swift Doorway app was generated largely through prompting, but CloudKit’s limited querying forced awkward manual checks and exposed many edge cases. Adam says the agents repeatedly find at least ten issues whenever asked for an adversarial review, raising the question of whether they are discovering meaningful flaws or simply generating more token-consuming criticism.

Adam explains the “gauntlet loop,” a workflow in which an orchestrator gives one agent the job of building and another fresh-context agent a demanding quality bar. The reviewer identifies failures, the builder fixes them, and the cycle repeats until the work passes; the builder must never also act as its own reviewer, even though the same model can fill both roles. Adam used this approach for an audio-editing task, limiting the number of loops because of cost. Robbie compares it with simply asking a model to think harder, but says his initial app was strong until layered features created brittleness.

The hosts endorse the distinction that “engineers aren’t being replaced, but coding is.” Robbie says AI can produce a large amount of code quickly, but it does not reliably make architectural decisions or account for performance, accessibility, edge cases, and maintainability. At Shopify, he sees pressure to ship immediately even when planning, ticketing, ownership, and estimation have not happened; he argues senior engineers should push back rather than accept arbitrary deadlines. Adam describes reviewing pull requests with a custom Claude prompt that looks for dead code and other cruft. A developer complained that his reviews slowed delivery, but Adam kept insisting that quality loops prevent future bugs.

Robbie and Adam distinguish a reviewer who blocks work for meaningful reasons from one who nitpicks irrelevant details. Adam appreciates colleagues who know a codebase deeply enough to recognize that changing one apparently harmless piece will break something elsewhere, even though their review temporarily blocks a pull request. He says teams need both quick reviews for small, safe changes and rigorous reviews for large or risky ones. The hosts agree that bugs can emerge from absurd sequences—such as rapidly hovering, escaping, and reopening a menu, or clicking something hundreds of times—but engineering judgment is still needed to decide which edge cases matter.

The hosts debate whether AI searches are more enclosed than conventional search engines, since neither queries the entire world and both optimize for useful-looking answers. Robbie finds AI more dishonest because it tries to please him, especially when he supplies leading hints; Adam notes that local models offer a less personalized alternative. Robbie’s local Muse Glimmer model works surprisingly well but responds slowly, prompting his Discord-based assistant “Punk Ass” to interrupt, mock Glimmer’s delay, and compete for attention. They also discuss invisible watermarks added to AI images and responses: someone open-sourced a tool stripping them the same day, making them useful for provenance checks but ineffective against deliberate misinformation.

Adam shows a Creator Micro, a six-button device customized to indicate which OpenCode terminal is idle, working, or waiting for input. He says the device turns agent supervision into a glanceable workflow. Robbie describes a more consequential example: he asked Punk Ass to discover the local API of his older Elegoo 3D printer, then create a private dashboard showing the printer’s video and remaining time without deploying anything publicly. The result ran on his home network and was made in minutes. They see this as personal, private software: the user specifies the outcome, while the agent handles the formerly tedious construction work.

Adam highlights two CSS developments. A scroll-axis property lets users move diagonally through interfaces with both horizontal and vertical scrolling instead of having a gesture lock into one direction; he says this could improve maps and large data tables. His second example, which he calls “CSS signals,” uses scroll-driven animation timelines and timeline scope to publish whether an element is entering or leaving a scrollport. Other elements can subscribe, enabling effects such as a sticky image changing in sync with section headers. He describes it as a lightweight CSS publish-subscribe mechanism that can express state without JavaScript intersection observers.

Robbie says the loss of coding’s old dopamine is real: he once took pride in recognizable CSS and implementation details, whereas now AI handles much of the bricklaying. He still wants generated software to be maintainable, but increasingly judges it by whether it solves the higher-level problem rather than whether it uses a particular framework. Adam recalls being told early in his career to study the Gang of Four design-pattern book; although its C++ examples were intimidating, he now values understanding what patterns accomplish and how they fit into an application. They argue that engineers can tell an AI which algorithm or pattern to use without manually implementing every detail.

Robbie reports that his robot mower, Jeff, cuts the grass effectively but repeatedly loses its position. Its RTK reference station needs a clear view of the sky, and weak connectivity near the house and surrounding trees causes Jeff to drift into the woods or even the street. When Jeff is out of bounds, Robbie cannot remotely command it to return, so a neighbor once moved it after finding it in the road while Robbie was away. Robbie thinks better cellular service would help, but says the mower should at least be able to move toward a known safe area. He closes by joining the Scotch Malt Whisky Society and inviting listeners to subscribe and review.