Tap Notes: Naming the Layer

The vocabulary is hardening. Simon Willison spent this week drawing a bright line between how professionals use agents and how everyone else does. Meanwhile, someone reverse-engineered Claude Code’s remote control architecture and found a bearer token functioning as a skeleton key. Anthropic shipped multi-session agent teams. WordPress quietly shipped an MCP adapter. These aren’t unrelated events — we’re in the naming-and-auditing phase of a new infrastructure layer, and the implementations aren’t waiting for the vocabulary to settle.

Writing About Agentic Engineering Patterns

Simon Willison draws the line between “vibe coding” (output, don’t look at the code) and “agentic engineering” (professional developers using agents to amplify existing expertise). The distinction sounds obvious once named. It wasn’t obvious before.

Why it matters: If vibe coding treats expertise as irrelevant, agentic engineering treats it as a multiplier. The skills you’ve spent years building don’t get replaced — they get leveraged at a different scale. That’s a different productivity thesis than “AI will replace developers,” and it’s the one that actually holds up when you’re debugging a broken agent in production at 2am. The naming matters because it gives practitioners a frame for what they’re doing that isn’t either breathless boosterism or existential dread.

Who Owns the Frontier

Anthropic shipped “Agent Teams” in Claude Code — multi-session task sharing, centralized management, inter-agent coordination as a first-party primitive.

Why it matters: Cross-interface session orchestration is becoming foundational. When Anthropic ships it as a platform feature days after independent builders are wiring it together by hand, the message is clear: the coordination primitives are being standardized. The real question isn’t whether to build on this layer — it’s whether your current agent setup can participate in it at all, or whether your agents are just maintaining separate context histories that can’t actually talk to each other. There’s a difference between “agents running” and “agents collaborating,” and the industry just named the gap.

Deep Dive: How Claude Code Remote Control Actually Works

AgentSteer dissects Claude Code’s remote control architecture. The session URL functions as an unbound bearer token — scan the QR code, get full access. The CLI polls; the remote sits blind for up to 10 minutes. No sequence numbers. No bidirectional health checks.

Why it matters: Convenience-first architectures defer the hard security questions. Here they come due. The specific failure — session URL as skeleton key — is one every agent infrastructure builder should recognize. If your agent handoff relies on “share this link,” you probably have a version of this problem. The one-way heartbeat is the less-discussed issue: the party that can’t initiate contact has no way to know whether it’s been orphaned or is just waiting. That ambiguity compounds in distributed systems, and calling something “working” doesn’t make it correct.

We Built an Open Source Tool That Lets You Click on UI Bugs and Have AI Agents Fix Them

An open-source MCP server that captures the exact DOM selector, computed styles, and bounding box when you click on a UI element, then hands that context to an agent for fixing. SQLite storage persists issues across sessions.

Why it matters: Describing visual glitches in prose — “the third card’s padding looks off” — is expensive and imprecise. You’re paying tokens for the agent to guess which div you mean. This solves the coordination problem: click the thing, get the structural context, let the agent work from facts. The SQLite persistence is the underrated part. Batching UI issues as a reviewable queue across sessions is a fundamentally different (better) workflow than trying to fix everything while the context is still warm.

From Abilities to AI Agents: Introducing the WordPress MCP Adapter

WordPress Core shipped an MCP Adapter that turns WP-CLI into an STDIO transport for AI agents. PHP plugins can register typed, permission-checked “Abilities” that any MCP client can discover and execute directly.

Why it matters: This is WordPress acknowledging that AI agents are clients, not just users. For the roughly 40% of the web running on WordPress, this is the bridge between “legacy PHP backend” and “AI-native tooling” — without rewriting anything. You register an Ability, the adapter handles the protocol translation. The pattern that gets adopted at scale is the one that meets the existing codebase where it lives. WordPress just made itself a viable AI-agent backend, which nobody saw coming three months ago.

2026.07: Aggregators and AI

Ben Thompson argues Spotify wins with AI not because it generates content, but because every user already has a unique experience in a monolithic network. AI becomes a sustaining technology for aggregators — not a disruptive threat. The $700B CapEx figure is the tell: infrastructure being locked in fast.

Why it matters: The reframe is useful and unsettling. If you’ve been thinking about AI primarily as a content creation tool, this analysis suggests you’re optimizing the wrong layer. Generation is commoditizing. The moat is in aggregation — personalized feeds creating lock-in that AI then enhances. Which means the window to establish the aggregation layer is a race against infrastructure consolidation, not an infinite runway. The companies building the personalization data structures right now are not going to be disrupted by cheaper generation. They’re going to use cheaper generation to deepen the moat they already have.

🪨