AI Synthesis Briefing: November 12, 2026
Bounded decision economics and proactive autonomous agents led intelligence developments today as Google DeepMind released the Gemini Decisions API evaluating 1,000 multimodal questions in 150ms at $0.05/M tokens with free outputs, OpenAI's Dots demonstrated background proactive monitoring across enterprise workspaces, DeepMind launched AlphaGenome 2 predicting DNA mutations across 107M individual cell profiles, and a16z mapped the emerging Business-to-Agent (B2A) economy.
Google Launches Gemini Decisions API: Sub-150ms Bounded Multimodal Decision Engine
Surfaced by The Deep View and TLDR Founders, Google DeepMind officially launched the Gemini Decisions API (gemini-decision-1), introducing a high-speed endpoint for bounded multimodal decision-making (tool selection, policy enforcement, model routing, and context grading). Priced at $0.10 per 2 million input tokens ($0.05/M) with free output tokens, the model evaluates up to 1,000 bounded questions in a single pass across text, images, video, and PDFs. With response latency as fast as 150 milliseconds (compared to 70–500ms for TypeSafe Jev and up to 329s for generative LLMs), Google reported internal deployments across Gemini Notebook routing and search MCP safety filters.
The Rise of Proactive Background Agents: OpenAI Dots & Multi-App Monitoring
Curated by The Deep View, enterprise reporting highlighted the transition from on-demand conversational assistants to persistent background agents, led by OpenAI Dots, Claude Duet, and Meta Muse. Operating within dedicated cloud sandbox environments, Dots continuously monitors connected email, Slack, calendar, and document streams to proactively catch travel booking date mismatches, unrecorded task commitments, and draft factual discrepancies prior to human publication. The architecture marks a structural shift toward autonomous background monitoring that preemptively assists operators while enforcing human sign-off on consequential mutations.
Google DeepMind AlphaGenome 2: Cellular-Resolution Disease & DNA Variant Mapping
Surfaced by The Deep View, The Neuron, and The Rundown AI, Google DeepMind unveiled AlphaGenome 2, a conditional DNA sequence model trained on over 107 million individual cell profiles across 68 human and mouse tissues. Overcoming the tissue-blur limitation of earlier models, AlphaGenome 2 operates at single-cell resolution, nearly doubling causal variant detection (30% vs. 16% on primary immune cells) and accurately predicting effect direction in 43% of cases (vs. 23%). Using conditional cell-state encoders, the model generalizes to unseen single-cell experiments without retraining and performs cross-modality imputation, predicting chromatin accessibility from gene expression.
The B2A (Business-to-Agent) Economy: Autonomous Procurement & Merchant Feeds
Curated by TLDR Founders and The Neuron, venture firm a16z published research detailing the emerging Business-to-Agent (B2A) economy, where commercial software and retail products reach customers through autonomous purchasing agents. In demonstration workflows, ChatGPT orchestrated multi-vendor video ad generation (calling Krea, Descript, and ElevenLabs) and converted paid software subscriptions autonomously. While airlines like United move to block bot bookings, Shopify reported orders originating from AI searches expanded nearly 13-fold in 12 months, driving merchants to deploy structured machine-readable catalog feeds.
Industry & Architecture Takeaways:
- Replace Text Generation with Bounded Decision Models: High-frequency agent steps (policy gating, tool routing, and triage verification) do not require verbose natural language output. Adopting non-generative decision models like Gemini Decisions API cuts routing latency to sub-150ms while driving evaluation costs down to $0.05 per million tokens with free outputs.
- Architect for Continuous Proactive Ingress: The evolution from user-prompted assistants to background proactive agents (OpenAI Dots, Claude Duet) requires continuous, asynchronous observation of user communication and calendar streams. Production architectures must incorporate persistent state listeners that cross-reference incoming confirmations against prior commitments without waiting for explicit chat triggers.
- Build B2A Machine-Readable Product Endpoints: As autonomous agents capture a growing share of consumer and enterprise purchasing (driving 13x annual order growth on Shopify), organizations must publish structured, machine-readable catalog schemas (
.well-known/ai-agent.json, API parameter endpoints) to ensure their products are discoverable and actionable by buyer agents.