Sectoral News/ signal over noise

The Model Context Protocol shipped in late 2024 as Anthropic's attempt to solve a very boring problem: every AI product was reinventing the wheel for tool calls, context injection, and data access. Different frameworks, different payload formats, the same pain. MCP proposes a single wire protocol so that any client — Claude, Cursor, custom agents — can talk to any server — a filesystem, a database, a CRM — without custom glue.

What changed in 2026 is adoption. Open-source MCP servers now exist for Postgres, Notion, GitHub, Linear, Slack, and dozens of smaller tools. Cursor and Claude Code ship with native MCP support. The result is something developers have been asking for since GPT-4: a plug-and-play tool ecosystem that doesn't require a new library every quarter.

The bet behind MCP is simple: the interface to AI models will matter less than the interface between AI models and the rest of your stack. If Anthropic is right, MCP will end up being the USB-C of the agent era — unglamorous, ubiquitous, and the thing everyone takes for granted two years from now.

For most of 2023 and 2024, "AI agent" was a demo video. Someone on Twitter would post a screen recording of an LLM booking a flight, the comments would argue about whether it was real, and the enterprise kept paying humans to do the same task. That window is closing. Agents that handle customer support triage, pull-request review, content moderation, and outbound sales are now in production at companies you've heard of.

The unlock wasn't bigger models — it was better orchestration. Frameworks like Claude Agent SDK, LangGraph, and vendor-specific offerings now give agents memory, tool access, and recovery paths that previous-generation prompts couldn't sustain. Reliability crossed an invisible threshold: when an agent fails nine times out of ten, you don't ship it; when it fails one in twenty, you start designing workflows around it.

The economic consequence is that the marginal cost of a knowledge worker is no longer tied to headcount in the same way. Small teams running agent fleets now rival mid-sized agencies in output. The companies adapting fastest aren't the biggest labs — they're small shops willing to rebuild their back office around agents instead of retrofitting agents onto org charts designed for humans.

Cloudflare Workers launched as a bet on V8 isolates — faster cold starts than containers, lower overhead than Lambda, global distribution by design. For years it was the "cool but niche" choice for edge logic. In 2026 the story is different. Workers, paired with Durable Objects, D1, R2, and Workers AI, is a full-stack platform that routinely handles production traffic for companies that would previously have defaulted to AWS without thinking.

What shifted is the surface area. Durable Objects solved the stateful-primitives problem that kept Workers out of real apps. R2 removed egress fees from object storage. Workers AI put LLM inference on the same runtime as your API. The single-platform pitch AWS used to win enterprise deals is now something Cloudflare can make with a straight face — and at a fraction of the operational cost.

For small and mid-sized teams the math is brutal. A full-stack Cloudflare deployment costs a fraction of the equivalent AWS bill, ships globally by default, and removes entire categories of infra work. Large enterprises still have lock-in and auditors to argue with, but every greenfield project in 2026 that doesn't at least consider Workers is leaving time and money on the table.

Astro's proposition was counterintuitive when it launched: ship zero JavaScript by default, hydrate only the interactive islands, treat content sites as content sites again. Three years later that proposition looks prescient. Marketing teams that fought with Next.js bundle sizes and hydration bugs are migrating to Astro in numbers big enough to show up in framework rankings.

The islands architecture turns out to be the right primitive for sites where 90% of the content is static and 10% needs interactivity. Your hero doesn't need React. Your nav probably doesn't either. Ship HTML for those and reserve the runtime cost for the one component that actually needs it. The result is Lighthouse scores that used to require heroic optimization effort, delivered by default.

Astro 5 and its content collections made the editorial story even tighter — Markdown-first, typed at build time, integrated with MDX and headless CMSs without the usual plumbing. For product teams still running Next on the marketing site because "we use React everywhere", the cost-benefit has flipped. Astro for content, Next or a framework of your choice for the app, and nobody has to apologize for the bundle size anymore.

Retrieval-Augmented Generation used to be a clever trick for grounding LLM output in custom data. In 2026 it's the default pattern. Every chatbot, every coding assistant, every internal Q&A tool is running semantic search across a private corpus before it writes a single token. The question stopped being "should we use RAG" and became "which vector database handles our scale".

The incumbents — Pinecone, Weaviate, Qdrant — are being squeezed by a new wave: Postgres with pgvector, SQLite with sqlite-vec, and turnkey options like Neon and Turso that put vector search where your data already lives. For most production workloads, keeping vectors next to rows beats operating a separate system. The performance gap has mostly closed, and the operational gap is enormous.

The interesting frontier is hybrid retrieval: combining keyword search, metadata filters, and vector similarity in a single query planner. Whoever ships the best developer ergonomics for hybrid search will own the next three years of AI-app infrastructure. If you're still pattern-matching on "vector DB vs document store" you're already behind.