The Bikeshed Pod is a weekly show where developers dive deep into the small but important details of software development that we all love to debate.
Reading the feed…
Scott, Dillon, and Matt dig into AI routers — the layer that decides which model actually answers your prompt — and why Stripe reportedly paid ~$9B for OpenRouter while Ramp built its own. The conversation covers what routers really do today (less 'smart judge,' more traffic control), why orgs want a single gateway for spend and model policy, and whether providers will quietly downgrade you to a cheaper model. Plus: burnout recovery, smart chess boards, flaky end-to-end tests, and a reindeer hot dog in Bergen.
Scott, Dillon, and Matt dig into AI agent skills — what they are, when they earn their place, and when a plain prompt does the same job. Scott's team routes every unit and integration test through a mandatory skill, Dillon dips in and out of them, and Matt loads almost none and lets the model figure it out. Plus anti-slop cleanups, the ponytail skill, handoff docs, and why ASD-STE100 is the only PR-description prompt Matt needs.
Scott, Dillon, and Matt dig into 'tokenomics' — the economics of AI models and whether the ever-climbing cost of running them is actually buying anyone more value. They swap notes on the rug-pull theory of AI pricing, tokenizer changes that quietly tax you 30%, and how their companies are clamping down (HubSpot defaulting to Sonnet and zeroing out the Fable budget). Plus a standup of weddings, World Cup corruption, a 16-0 soccer loss, and a first-ever Fourth of July party.
Dillon, Scott, and Matt dig into automated testing — the high upfront cost of getting started, why smoke tests deserve to gate code before it hits master, and whether enforced page object models are a useful abstraction or just a layer of indirection. Scott makes the case that smoke tests should come first while his coworkers insist the end-to-end suite will catch everything, and the crew lands on a surprising amount of agreement for a show built on disagreement. Plus: a launch that turned into permanent backend ownership, a 75-day estimate squeezed into two months, and a CLI feedback hook for
Dillon, Scott, and Matt dust off the parking lot for the first time in seven months to work through a backlog of tech news. They open with the 'everything app' land grab — Google's new Whoop competitor, Uber muscling onto Airbnb's turf — and where the line sits between a smart adjacent bet and plain enshittification. From there it's the Cloudflare and Coinbase layoffs and whether 'because AI' actually holds up, then RoboBun quietly becoming Bun's top committer and out-pacing the human who's worked on it for four years. Plus Pi the 'Vim of agent harnesses,' tech bros flexing token spend, and th
Dillon, Scott, and Matt dig into a Harvard Business Review piece on the central irony of the AI era: the tools sold to reduce our workload are quietly making us do more, because it's cheap to start work and expensive to finish it. They get into agent-swarm mandates and the dreaded 'tech lead of agents' adoption tier, Dillon's ever-growing pile of 50 open PRs, skill atrophy, and companies torching their entire AI budgets in a quarter. Plus a brand-new drama segment — Spill the Sweet Tea — on TanStack Start's controversial 'RSC support' and whether it actually follows the React spec.
Dillon, Scott, and Matt dig into Matt's hot take that you shouldn't be using plan mode in your AI coding agents anymore. They land on a nuanced agreement — planning still matters, but plan mode in Claude Code and Codex has stopped pulling its weight, spinning for 30 minutes and dumping a markdown doc instead of asking the clarifying questions it used to. Plus grill-me skills, POC-first workflows, Scott's eighth powerlifting meet, and Matt accidentally reinventing Tanstack Query.
Dillon walks Scott and Matt through the daily briefing dashboard he's been vibe-coding at work — a single page that aggregates his to-dos, PRs, JIRA tickets, Datadog alerts, and recent notes — and shares the custom Claude skills he's built around it. The conversation explores AI as a fast track to personal software, where AI agents struggle (keeping consistent UI), where they shine (generating visual diagrams to understand problems), and the trap of suddenly having 20 open PRs because you can ship faster than you can review.
Dillon, Scott, and Matt dig into the drama around Matteo Collina's 19,000-line Node.js PR — written largely with Claude Code over Christmas break — that's forced the Node.js community to confront how AI-assisted contributions fit into open source governance, the DCO, and the future of runtimes like Bun and Deno.
Dillon, Scott, and Matt dig into what happens when PR feedback drifts from the plan the team thought they aligned on — scope creep, nitpick comments that block merges, and the eternal tension between shipping the incremental thing and building the "ideal" thing. They land on communication, lightweight design docs, and a healthier code review culture as the way through. Plus: Dillon discovers token anxiety after burning $300 of Opus in four hours, and Scott spoils Project Hail Mary for Matt.
Dillon, Scott, and Matt unpack the chaotic saga of ClawdBot → MoldBot → OpenClaw — the rapidly-rebranded autonomous agent platform that's exploded on GitHub — and dig into the marketplace malware scandal, the credentials-access risks, and what Anthropic cutting off subscription billing tells us about the economics of agentic AI.
Matt, Scott, and Dillon push back on the AI hype cycle by digging into a Dax tweet about what AI adoption actually looks like inside larger tech companies — bloated LLM bills, unmotivated 9-to-5ers churning out slop, and the burned-out few left cleaning it all up. Plus thoughts on whether AI is really just a fancy search engine, and a hot take that the Super Bowl MVP should've gone to the kicker.
Dillon, Scott, and Matt discuss how to avoid the 'slop pit' of AI-assisted development — the tendency to accept mediocre AI-generated code without proper scrutiny. They share practical strategies for maintaining code quality including upfront planning workflows, multi-layered code review with AI and human reviewers, TDD-inspired approaches to AI coding, and the role of linting and type safety as guardrails.
Scott, Dillon, and Matt share their boldest (and not-so-bold) predictions for 2026 — from autonomous coding agents replacing engineers, to the AI bubble bursting, to one of the hosts getting laid off. They rate each prediction on a Taco Bell spice scale and wrap up with standup updates on AI tool usage, honeymoon planning, and the eternal struggle of AI-generated code quality.
The bikeshed boys discuss two major tech industry developments: Anthropic's crackdown on third-party agent harnesses and Tailwind's recent layoffs. They explore how Anthropic is shifting focus from the model itself to their suite of tools to create stickiness, and how AI tools like Claude are disrupting the component library business model that Tailwind relied on.
It's review season, and the bikeshed boys dive into the dreaded annual performance review process—and how AI is changing everything about it. They share strategies for writing effective self-reviews, debate whether managers already know where they'll rank you before reading a word, and explore the messy dynamics of team politics: fighting for interesting projects, dealing with too many senior engineers, and avoiding the "senior trap" of over-architecting instead of shipping. Plus: New Year's resolutions, Claude Max vs. Cursor debates, and why Arc Raiders lobbies are way sweatier in trios.
This week on The Bikeshed, Scott, Matt, and Dillon tackle two breaking stories that have the JavaScript community buzzing: React's severity 10 vulnerability in React Server Components and Anthropic's surprising acquisition of Bun.
The bikeshed boys dive deep into the emerging world of AI-powered code reviews—exploring whether our robot overlords are ready to stamp your PRs, or if they're just glorified linters with delusions of grandeur.
The bikeshed boys tackle one of software engineering's most contentious debates: should you build internal tools or adopt external solutions? With real examples from their companies—custom dependency managers, GitHub wrappers, and developer portals—they explore when it makes sense to roll your own versus when you're just creating expensive technical debt. Expect war stories, a framework for build-vs-buy decisions, and Dillon's first truly spicy take about Vercel.
The bikeshed boys ramble on about coffee, Cursor 2.0 and their new composer model, Claude Code tips and tricks, and pitfalls, as well as AI in coding interviews!
The bikeshed boys talk about the new hottest web framework on the block: Remix 3, hopefully its third times the charm for the remix guys!
AI crawlers killing ads, chatbots replacing browsers, and the slow death of human-centric content. The bikeshed boys debate whether we're heading for a hard fork of the web or just a really awkward transition.
The bikeshed boys talk about whether chasing 5-minute deploys is worth it, debating the trade-offs between deployment speed and stability, and why moving fast might sometimes mask the real problems slowing your team down.
The bikeshed boys share how FAANG-like companies manage their frontend dependencies without pain! Hint! They Don't!
The bikeshed boys interview Matt on this episode, talking about all things React Server Components (RSCs), exploring their benefits, trade-offs, and adoption strategies. The hosts draw on their experiences at Wayfair, Fireworks, and Whoop to provide practical insights for learning and adopting RSCs.
