Three things happened in roughly thirty-six hours this week, and they all point at the same place.

On Tuesday, Andrej Karpathy — the researcher who in November 2025 published a few-hundred-line web app called llm-council and accidentally seeded the entire post-2025 council ecosystem — announced he's joining Anthropic's pre-training team under Nick Joseph. His stated focus: using Claude to accelerate pre-training research. He started this week.

On Tuesday into Wednesday, Google ran I/O 2026 and shipped Antigravity 2.0 — a standalone desktop agent-orchestration platform with a CLI, an SDK, Managed Agents in the Gemini API, and Enterprise support through the new Gemini Enterprise Agent Platform. The keynote demo was the part the developer feeds will remember: 93 sub-agents in parallel, 2.6 billion tokens, the framework of an operating system built in about twelve hours. All running on Gemini 3.5 Flash, which Google says outperforms Gemini 3.1 Pro across almost all benchmarks while running four times faster.

On Tuesday, KPMG and Anthropic signed a global alliance and launched KPMG Digital Gateway Powered by Claude. KPMG's 276,000-person workforce gets Claude across the Microsoft Azure–built Digital Gateway. Initial focus: tax clients and private equity firms. Claude Code embedded into KPMG Blaze for IT modernization. Co-developed Claude-powered products for PE portfolio companies.

Read those three together and you get a single architectural argument. The substrate layer has two planes now. The architecture layer has six primitives. The strategy layer is where product surfaces fragment. And underneath everything else, pre-training of the Chairman model is the layer that doesn't commoditize.

Substrate has two planes

Yesterday we wrote about the substrate layer consolidating across hyperscalers via open standards — Anthropic Agent Skills as the open standard (six-vendor adoption: Microsoft VS Code + GitHub Copilot, OpenAI ChatGPT + Codex CLI, Atlassian, Figma, Cursor), MCP at 110M+ monthly SDK downloads, the Agentic AI Foundation at 170+ Linux Foundation member organizations in under four months, and the Code as Agent Harness survey at arXiv:2605.18747 as the academic synthesis.

Antigravity 2.0 forces a refinement. Substrate has two planes.

Standards plane: MCP, Agent Skills, AGENTS.md, Code as Agent Harness. This is where the hyperscalers cooperate, via AAIF and the Linux Foundation. It looks like commodity infrastructure because it is.

Runtime plane: Anthropic Project Glasswing. OpenAI Codex Security. Claude Managed Agents. LangSmith Deployment. And now, with Antigravity 2.0 plus the Gemini Enterprise Agent Platform, Google. This is where the hyperscalers compete — on CLIs, SDKs, managed execution, developer ergonomics, and enterprise pricing.

The substrate is commodity at the interface. Competitive at the implementation. Both things are true at the same time, on what looks like the same layer.

Architecture now has six primitives

We tracked five fixes yesterday — heterogeneity (Council Mode, arXiv:2604.02923), hierarchy (CHAL, arXiv:2605.12718), role separation (the Glasswing + Daybreak + MDASH production pattern), adaptive gating (Cost of Consensus II, arXiv:2605.06988), and arena-based argumentation with provenance (Contestable MAD, arXiv:2605.14495). Each fix mapped to a named failure mode of flat debate.

A sixth primitive landed quietly this week. arXiv:2605.08268 — Insider Attacks in Multi-Agent LLM Consensus Systems (Sun, Liu, Hu, Zheng) — formalizes a malicious agent participating as a legitimate group member while pursuing a hidden adversarial goal of delaying or preventing agreement. It's a world-model-based attacker framework trained with reinforcement learning. The structural concern has a name now: adversary-resistant consensus.

That makes the running taxonomy six wide, across two methodological dimensions. Five epistemic fixes for five epistemic failure modes. One security primitive for adversarial deployment. Six primitives, six failure-or-attack modes, all named in the April-to-May 2026 window.

Strategy fragments into product surfaces

Apple at WWDC 2026 (eighteen days out, June 8) routes one model — Siri picks, then forwards. Perplexity Model Council ships one default deliberation pattern (GPT 5.4 + Claude Opus 4.6 + Gemini 3.1 Pro, with Claude as the synthesizer) locked to consumer Max. Microsoft 365 Copilot Critique + Council ships GPT-drafts-Claude-reviews as one product surface. Google Antigravity 2.0 ships one orchestration default with Gemini 3.5 Flash as the homogeneous worker pool. The hyperscaler verticals (Glasswing for security, Anthropic's Financial Services push, KPMG Digital Gateway for tax and PE) each ship one vertical-specific managed package.

A small thing to notice about Antigravity's 93-sub-agent demo: 93 Gemini 3.5 Flash workers in parallel is structurally a Traditional Council or Survivor configuration on a homogeneous worker pool. The Cost of Consensus paper (arXiv:2605.00914, Bertalanič & Fortuna at the Jožef Stefan Institute) measured 10-agent homogeneous teams: 85.5% sycophantic conformity, 70% contextual fragility, 2.1–3.4× token cost, and a single 8B model self-correcting beats the team. Whether 93 at hyperscaler scale resists those modes is the empirical question the next paper-wave is going to answer. The current answer in the literature is: only with heterogeneity, hierarchy, role separation, and adaptive gating.

Shingikai ships six strategies — Traditional Council, Round Robin, Survivor, Collaborative Editing, Red Team vs. Blue Team, and Quick Take. The user picks which pattern the question warrants. 200+ worker models, user-selectable Chairman. The strategy layer is where the explicit engineering choice lives once substrate consolidates and architecture stabilizes.

Pre-training is the layer underneath

The Karpathy hire is the part that's structurally hardest to overstate. Claude is the Chairman model in Perplexity Model Council. It's the reviewer in Microsoft Critique + Council. It's the auditor in Glasswing. It's the Claude in KPMG's Digital Gateway across 276,000 seats — that alliance signed the same day Karpathy started. It's the default Chairman for Shingikai.

The Chairman position is the highest-leverage architectural surface in any council deployment. It's what preserves disagreement honestly, surfaces support-and-attack reasoning, decides what to compress into the synthesis, and resists adversarial manipulation by an aligned-looking participant. The five named failure modes describe what happens when the Chairman isn't doing its job. The six primitives describe what it's supposed to do.

And the researcher who named the council-product pattern in late 2025 is now at the company that makes the model that runs that role across the industry, doing pre-training research on it.

The originator of the pattern is optimizing the model that runs the pattern.

Where this leaves us

Substrate is the Linux Foundation's job at the standards plane and the hyperscalers' job at the runtime plane. Architecture is the academic literature's job — six primitives, six failure-or-attack modes, six anchor papers in four weeks. Pre-training is Anthropic's job — Karpathy's now there. Strategy is the user's job — pick the pattern that matches the question.

Six strategies, because the literature now names six primitives, and the Chairman model that runs the synthesis layer is being improved as we type.

shingik.ai — heterogeneous frontier council, six explicit strategies, no signup.