168: Are Dev Conferences Dying?

Jack reports that React Advanced, a Git Nation conference held at London’s Brewery venue, was noticeably smaller than his previous appearances there. A main room that could normally hold 400–500 people had roughly 200–300 attendees, with more widely spaced seating and no overflow area. He says the decline was not limited to his talk and reflects a broader trend: coding conferences are reportedly down about 50%, while AI conferences are thriving. Jack also did interviews for future podcasts and plans to revise his AI-focused talk after seeing attendees leave during it.

The hosts debate whether React’s decline is part of the normal lifecycle that Jack saw during the fall of jQuery, or whether AI has accelerated it across front-end topics. Jack argues that agents now handle much of the procedural, step-by-step coding work, so developers are more interested in architecture, API contracts, orchestration, and strategy. He cites conference conversations about Effect and the new YAK CSS-in-JS framework as evidence that conventional topics still exist, but says the strongest demand is now for practical ways to organize agents and workflows.

Paige notes that explaining agent-driven work is difficult because each developer has taught an agent different preferences and quirks; a workflow that works for one person may not transfer directly to another. The hosts suggest that reusable skills and workflows can instead be shared online and adapted by agents. Jack highlights Matt Pocock’s one-way-door/two-way-door framework for code reviews: an irreversible database change deserves scrutiny, while a reversible CSS adjustment may be safe to merge automatically. They argue conferences still provide value when they expose useful decision-making patterns rather than merely showing code.

TJ summarizes Addy Osmani’s account of how the Claude Code team handles its monorepo as agents increase code volume. The team went from approximately 2,500 merged changes per day in May to more than 30,000 in September, making human review of every change impossible. Every pull request still receives standard CI checks such as linting and type checking, but agents also hunt for bugs and perform dedicated security passes. Pull requests are assigned low-, medium-, or high-risk tiers based on the change itself, not on who authored it.

Claude Code’s low-risk tier includes documentation, copy changes, internal-only features, and additive changes protected by feature flags; those changes receive automated checks but no human review. Ordinary product changes and sensitive areas such as caching, the auto-updater, and system prompts require one human reviewer. High-risk paths, including authentication and billing, require two human approvals. The team relies heavily on feature flags and internal-first releases, so merged changes can be exercised internally before public exposure. When a production bug slips through, an incident review asks what automated check or process could have caught it.

TJ says the most immediately useful idea is allowing routine dependency updates and similarly harmless pull requests to merge automatically after monitoring whether a triage bot classifies them correctly. Paige agrees that internal documentation or mundane release notes may be safe to automate, but warns that external documentation—especially tutorials and instructions—can look correct to an agent while remaining confusing to readers. The group maps the one-reviewer and two-reviewer tiers to reversible and irreversible changes, while imagining an assembly line in which triage, coding, testing, and design agents each handle an appropriate part of the process.

The hosts describe GitHub Universe’s 2026 announcements as heavily focused on agents, while also responding to reliability concerns. The Copilot app is a downloadable desktop environment for coding tasks, similar in broad form to other agent applications. Copilot Fusion is intended to select an appropriate model for reasoning, code generation, debugging, and tool use, avoiding the expense of using a frontier model for trivial work. Paige and TJ see the enterprise appeal—Microsoft can offer budgets, reporting, and administrative controls—but prefer retaining direct control over which model handles important tasks.

Copilot Projects is designed to pursue a larger goal across multiple repositories, from ideation and experimentation through implementation, using longer-running loops. TJ compares this with workflows where a powerful model coordinates several repositories but wastes expensive tokens on simple edits; he would prefer the system to delegate easy tasks to cheaper subagents while reserving advanced models for difficult coordination. GitHub also announced session continuity across the Copilot app, VS Code, the CLI, and web interfaces, including the ability to continue or prompt work from a phone. Other features include repository automations and a deeper Copilot Code Review Max mode.

GitHub says it added 4 million CPU cores, 160 petabytes of high-speed storage, and 50 terabytes per second of network capacity over the past year, while migrating toward a multi-region Microsoft Azure deployment. The hosts interpret this as an attempt to address reliability problems without explicitly discussing recent outages. They also question whether GitHub’s Git-based foundation can sustain agent-driven development at extreme volume. Jack expects Git to remain the underlying substrate, though some GitHub abstractions may change; TJ says his own projects are nowhere near the scale at which Git becomes a bottleneck.

TJ celebrates Jeff Graham purchasing CSS-Tricks from DigitalOcean after working on the site for about 12 years. DigitalOcean had initially produced content after acquiring the site from Chris Coyer in 2022, but the project later languished. Graham plans to support it through advertising and memberships while acknowledging that educational content has been struggling. The hosts hope CSS-Tricks can adapt beyond its original focus on small CSS tricks. They then discuss Arc’s announcement that the browser will sunset on December 22, 2026, with users encouraged to move to Dia; Atlassian reportedly acquired its parent company for $610 million in cash.

Paige recommends Netflix’s The Diplomat, praising its tense, intelligent political story about a U.S. ambassador in London and her politically involved husband; she warns it requires close attention. TJ recommends Stuart Turton’s The 7½ Deaths of Evelyn Hardcastle, a murder mystery with a Groundhog Day-like repeated day whose details must be tracked carefully. Jack recommends the Dungeon Crawler Carl series, especially its audiobooks, arguing that its apparently silly alien-dungeon premise develops deep character arcs and long-running payoffs. He also recommends the stop-motion film Wildwood, made by Laika and set partly in a fantastical version of Portland.