How 12-Factor Agents Solve the 80% Quality Wall in Production

Anthropic overtakes OpenAI in revenue run rate as Microsoft pivots to internal models

Anthropic's annualized revenue run rate has reached $47 billion, surpassing OpenAI's most recent disclosure of $25–33 billion, according to Deutsche Bank Research analyst Adrian Cox citing Ramp payment data. Similarweb data shows ChatGPT monthly visits fell below majority market share for the first time. Both companies view this metric as central to their IPO narratives—Anthropic targeting October, OpenAI September. Simultaneously, Microsoft is deploying internally-built MAI models across Excel and Outlook to replace Anthropic and OpenAI, with stated goal to eliminate external model spending. Microsoft launched seven MAI variants at Build conference including one matching Anthropic's Opus 4.6 at lower cost, demonstrating how distribution partners can rapidly become competitors and directly threatening the API-dependent business models of frontier labs.

JADEPUFFER: First autonomous AI ransomware attack executed 600+ payloads without human direction

Sysdig's Threat Research Team documented JADEPUFFER, the first end-to-end autonomous AI ransomware attack where an LLM agent independently conducted reconnaissance, credential harvesting, lateral movement, privilege escalation, persistence, database encryption, data destruction, and ransom note generation—executing more than 600 distinct payloads in a compressed time window. A human operator only selected the initial target and set up infrastructure; the agent self-narrated every action in natural-language code comments and self-corrected errors in real time. Sysdig could not identify which specific LLM powered the agent; API keys for OpenAI, Anthropic, DeepSeek, and Gemini found in logs were credentials the agent stole during compromise, not models driving the attack. The incident answers a critical question: whether complex multi-stage attacks can execute autonomously without human direction at each step. The answer is yes.

Thinking Machines releases Inkling: 975B-parameter open-weight MoE model ranks 9th on agentic benchmarks

Thinking Machines Lab released Inkling, a 975B-parameter open-weights Mixture-of-Experts model with 41B active parameters, trained from scratch on 45 trillion tokens of text, images, audio, and video. The model supports a 1M-token context window, features controllable thinking effort for performance-token tradeoffs, and achieved an Elo of 1257 on Design Arena's Agentic Web App Arena—ranking 9th overall and highest among US-based open-weight models for agentic workloads, placing it in the same performance band as Claude Opus 4.6 and Gemini 3.5 Flash. Inkling ships with day-one support across vLLM, SGLang, Modal, and other major inference stacks under Apache 2.0 licensing, with a smaller 12B-parameter variant also previewed. The release signals a strategic bet on customizable foundation models over benchmark maximization and represents the strongest US-based open-weight model release to date.

Demis Hassabis proposes U.S. Frontier AI Standards Body modeled on FINRA

Demis Hassabis outlined a federally overseen public-private partnership to develop dynamic benchmarks, conduct pre-release safety evaluations, and coordinate international standards for frontier-class AI models. The proposal emphasizes scientific rigor and adaptability over static rules, with industry funding and potential mandatory participation, positioning it as essential governance for AGI development arriving within years.

China's Q2 GDP growth misses at 4.3% as property investment falls 18% while AI and high-tech manufacturing expand

China's Q2 real GDP growth declined to 4.3% year-on-year, below the 4.5% consensus and the lower bound of the government's 4.5–5.0% full-year target. Property investment fell 18% year-on-year while AI and high-tech manufacturing expanded rapidly. The GDP deflator turned positive for the first time in 12 quarters, ending China's three-year deflationary streak. Officials noted that "new growth drivers" contributed more than 40% of economic growth, signaling structural economic weakness in traditional sectors offset by strength in emerging tech industries—relevant to understanding China's AI infrastructure investment priorities.

12-Factor Agents

Dex Horthy in GitHub / HumanLayer published an essay titled "12-Factor Agents". Main takeaways:

• Most agent frameworks fail in production because teams hit 70-80% quality and realize that getting past 80% requires reverse-engineering the framework; the fastest path to high-quality customer-facing agents is incorporating modular agent concepts into existing products rather than greenfield framework adoption.
• The core agent loop (LLM determines next step → deterministic execution → append result to context → repeat) doesn't work well in practice; successful agents are mostly deterministic software with LLM steps at strategic points, not pure agentic loops.
• Context engineering (owning prompts, context windows, control flow, error handling, and execution state) is the critical differentiator; the 12 factors provide a systematic approach to managing these surfaces across tool calls, state management, human contact points, and stateless reducer patterns.

Builder-Deployed Markets Overtake Crypto on Hyperliquid

The Defiant reported on a structural shift in Hyperliquid's trading composition, where builder-deployed markets for stocks, commodities, and indices have begun outtrading native crypto perpetuals. Main takeaways:

• Builder markets generated $5.41B (51.8%) of Hyperliquid's $10.44B total notional volume on July 15, exceeding native crypto perps' $5.03B for the first time on weekdays
• Builder-market share climbed from near-zero at HIP-3 launch (October 2025) to 50%+ by July 2026, peaking on weekdays while crypto dominates weekends
• Single builder (trade.xyz) operates all but a fraction of HIP-3 volume; SK Hynix alone accounts for $1.62B of equity volume, suggesting concentration risk
• Hyperliquid settles ~70% of onchain perpetual futures volume with $840M annualized revenue run rate, positioning it as potential 24/7 trading layer for traditional assets

Thinking Machines Lab Debuts Inkling: 975B Parameter Open-Weight Multimodal MoE Model with Controllable Thinking Effort

Thinking Machines Lab released Inkling, a 975B-parameter open-weight mixture-of-experts model with 41B active parameters trained on 45 trillion tokens. Main takeaways:

• Inkling supports 1M token context, native multimodal (text/image/audio), and controllable thinking effort with variable token efficiency across benchmarks
• Achieves 1/3 token efficiency vs. Nemotron 3 Ultra on Terminal Bench 2.1 agentic coding; scores competitively on forecasting (61.1 Brier Index), vision (73.5% MMMU Pro), and audio (91.4% VoiceBench)
• Trained for epistemics (calibration, instruction-following, censorship resistance) and safety; exhibits 78% adversarial robustness on FORTRESS and 98.6% refusal rate on StrongREJECT
• Available for fine-tuning on Tinker platform with Inkling Playground for interactive testing; includes Inkling-Small variant with 12B active parameters for lower-cost deployment

Thinking Machines Releases Inkling: 975B-Parameter Open-Weight Multimodal Model with 1M Context

Thinking Machines Lab in Thinking Machines / Hugging Face published a release titled "Thinking Machines Releases Inkling: 975B-Parameter Open-Weight Multimodal Model with 1M Context". Main takeaways:

• First large-scale open-weight model natively supporting text, image, and audio inputs with 1M context window and Apache 2.0 licensing; includes Inkling-Small variant (276B total / 12B active)
• Architecture innovations: relative positional attention (vs. RoPE), hybrid sliding-window/global attention (5:1 ratio), short convolution layers, MoE with 2 shared experts, and speculative decoding via 8 MTP heads
• Ranks #41 on Artificial Analysis Intelligence Index (highest U.S. open-weight model), ahead of Nemotron 3 Ultra (38); #9 on Design Arena Agentic Web Dev leaderboard; token-efficient at 25K output tokens per task vs. 37-43K for competitors
• Day-0 ecosystem support across vLLM (380 tok/s on 4× GB200 with MTP), SGLang, Modal (DFlash speculator for 67% throughput gain), Baseten, Databricks; Tinker API pricing $1.87-$3.74 per 1M input tokens depending on context window
• Trained on GB300 NVL72; NVFP4 quantized checkpoint available; Unsloth released 1-bit GGUF quants at 86% size reduction (270GB vs 1.9TB) with 74.2% accuracy retention