MCP and A2A Are Becoming the Plumbing of the Agent Economy

MCP and A2A Are Becoming the Plumbing of the Agent Economy

Two protocols are quietly becoming as important as the models themselves. MCP (Anthropic, Nov 2024, now under Linux Foundation/AAIF) standardizes how agents talk to tools and hit 97 million monthly SDK downloads by March 2026 [1]. A2A (Google, Apr 2025, also Linux Foundation) handles the harder problem — agent discovery, negotiation, and multi-turn collaboration between agents via Agent Cards and JSON-RPC [3].

AWS just expanded A2A support in Amazon Connect for voice and text delegation, with unified observability and guardrails baked in [1]. That's not a toy demo — that's enterprise contact-center infrastructure betting on agents delegating to other agents as a default pattern, not an edge case.

This is the unglamorous infrastructure layer that determines whether "multi-agent systems" are a real production pattern or a conference-slide fantasy. Builders on X have been trading war stories about sandboxing agent execution safely and watching agents attempt creative workarounds — the engineering reality of orchestration is messier than the pitch decks suggest, and that's exactly where the real work is happening right now.

Vibe Coding Goes Mainstream — And It's Not Just a Meme Anymore

"Vibe coding" — describing intent in natural language and letting AI agents handle implementation — has gone from Twitter joke to dominant paradigm faster than almost anyone predicted. GitHub reports roughly 46% of new code is now AI-generated, Gartner projects 60% by year-end, and about a quarter of YC startups are shipping mostly AI-generated codebases [1][2].

Tools like Claude Code, Lovable, Bolt, and Cursor are the engines behind this, and the shift is real enough that it's forcing a rethink of what "building software" even means at a startup [3]. The conversation on X has moved past "does this work" to "what does my job look like when the AI writes the implementation and I own the judgment calls."

That's the whole thesis of this newsletter, honestly. Code is free now. The scarce skill is knowing what to build, why, and when to say no to the AI's first draft.

Europe Wakes Up to Its Frontier AI Gap

EU AI Act enforcement kicked in fully on August 2, 2026, with the AI Office and national authorities now overseeing systemic risk for frontier models, backed by fines up to €15M or 3% of global turnover [2]. But the more interesting document this month is the July Expert Forum on Frontier AI report, which bluntly states Europe has no frontier labs competing at the pace of the US or China, and calls for urgent action on compute, capital, talent, and data [1].

The Apply AI Strategy now includes a "frontier grand challenge," and scientific panel rules were formally adopted in September — Europe is building governance infrastructure faster than it's building labs [1][2]. That asymmetry is the whole story: the EU is world-class at regulating a technology it isn't yet producing at the frontier.

For Nordic companies specifically, this cuts both ways. Regulatory clarity is a genuine advantage for enterprise adoption — customers trust compliant systems. But it also means the best models your team will use in 2027 are still very likely coming from Anthropic, OpenAI, or Chinese labs, not from anything built in Brussels or Stockholm.