AI AgentSpecialized Decisions, Shared Work
New models separate decisions, research, and generation. Collaboration tools turn discussion into usable interfaces, while local tools bring fresh resource tradeoffs.
The RISC Machine
RISC = the four systems of a production-grade agent or robot
A production-grade agent needs more than a brain. It must keep operating, survive failure, resist overreach, and participate in real organizational workflows.
AI Agent Radar
Specialist Models Narrow the Output
Decision returns probabilities; AlphaGenome 2 predicts effects under a cell condition. Narrowly defined outputs make the right checks easier to specify.
Collaboration Tools Preserve Next Steps
Tabletop turns discussion into shared controls; ElevenAI turns meetings into action items. Both depend on whether participants can actually continue the work.
Local Tools Shift Resource Tradeoffs
Hash LMG keeps production on a computer; tsfm-pg puts inference inside a database. Less data movement still requires checking coexistence with the host workload.
Notable New Capabilities
Gemini Decision Separates Classification From Generation
What happened: Google released Gemini Decision and the Decisions API on November 11. The specialized model returns probabilities over text, images, audio, video, and PDFs without generating output text. One request can evaluate up to 1,000 independent yes/no, single-choice, or rubric-rating questions. The shared input and all questions must fit within a 262,144-token context window.
Why it matters: Routing and scoring need not run through a text-generation loop. A dedicated endpoint makes latency, calibration, and business thresholds distinct engineering decisions.
Who should care: Teams building multimodal classifiers, support routing, policy checks, and agent tool selection.
Qwen4 Preview Is Limited to Token Plan Subscribers
What happened: QwenCloud introduced Qwen4-Max-Preview on November 11, combining text, image, and video understanding with tool calling and structured outputs. It lists a 1M-token context and a 393,216-token output budget shared by reasoning and the final answer. Access is limited to Token Plan; pay-as-you-go is unavailable.
Why it matters: This is a window for multimodal evaluation and migration testing, not a stable production commitment: the provider explicitly schedules retirement after the preview.
Who should care: Token Plan developers and teams testing long-context coding or visual agents.
AgentCore Exposes Web Search as a Standalone API
What happened: AWS introduced a direct Web Search API for AgentCore, without first creating a Gateway or MCP target. SDK calls query Amazon’s own index and return ranked results with text extracts. Domain and publication-date filters are supported. Authentication uses IAM and SigV4, with optional PrivateLink connectivity.
Why it matters: Search no longer requires adopting an entire agent gateway. Existing orchestration and data pipelines can integrate it directly, while Gateway remains an option for shared policies and multiple tools.
Who should care: Teams with AWS applications, custom agent loops, or fact-checking pipelines that need web search.
Tabletop Turns Group Plans into Shared Interactive Tools
What happened: Google Labs introduced Tabletop, a shared space that turns links, files and prompts into interactive itineraries, polls, expense splitters and tournament brackets. Participants can use the same space together through an invitation link.
Why it matters: It turns a conversation into an interface the group can operate together, rather than another summary to copy into a spreadsheet.
Who should care: People organizing family trips, recreational leagues or group meals.
Brief: ElevenAI Connects Dictation with Meeting Follow-through
What happened: At its New York Summit, ElevenLabs previewed ElevenAI, using Scribe v3 for dictation and meeting notes. With participant consent, it produces summaries and attendee action items. The company says free early access opens that day.
Why it matters: The useful step beyond transcription is preserving who needs to do what after a conversation.
Who should care: Enterprise teams handling interviews and everyday meetings.
Brief: Zumungo Turns Home Automation Requests into Reviewable Rules
What happened: Zumungo runs on a home computer and lets users draw rooms, discover devices and describe automation events or control panels in natural language. Its website says generated events are shown before confirmation, with optional cloud AI or local AI on an NVIDIA system with at least 16 GB VRAM.
Why it matters: It aims to simplify cross-brand configuration while separating generated rules from approval to save them.
Who should care: Households already using smart-home devices across several apps.
GitHub Discoveries
Stars are an attention snapshot at publication time, not a quality ranking.
Brief: Run Time-Series Forecasts Inside Postgres
What happened: tsfm-pg runs Chronos-2 and Timestral inside a Postgres extension, returning forecast quantiles through SQL. Models stay on the database host rather than sending series to a cloud API. The author showcased it today.
Why it matters: Forecasting stays close to stored data, but shares database compute resources.
Who should care: Teams storing metrics or business time series in Postgres.
Hash LMG Lets Small Models Storyboard While the App Builds the Film
What happened: Open Hash LMG splits production between a model and an application. A small local model writes a storyboard and simple drawings; the app handles layout, timing, motion and sound, with checks for text size, overflow and reading time. Its README links Apple Silicon downloads and GPL source, and lists export formats including MP4.
Why it matters: Rule-based layout and rendering reduce what a small model must solve, while leaving a timeline for manual editing.
Who should care: Independent creators making product shorts and kinetic typography locally.
Global Technology and Market Watch
AlphaGenome 2 Conditions Variant Predictions on Cell State
What happened: DeepMind announced AlphaGenome 2 on November 11. It combines a million-letter DNA sequence with a specified cell type or state to predict variant effects on gene regulation. The model was trained on more than 100 million cell profiles. It is available through the AlphaGenome API for noncommercial research, with commercial research in private preview.
Why it matters: Moving beyond tissue averages can expose regulatory effects that appear only in particular immune states or disease-relevant cells.
Who should care: Researchers studying genetic variation, single-cell biology, and disease mechanisms.
Meta Agents Refine Compute Allocation
What happened: Meta described how Intelligent Delivery allocates compute across regions, hardware, and ranking stages based on expected value per request. Agents read source code, design documents, and years of experiment logs to propose changes and design experiments. Engineers approve deployment. Meta says agents now contribute about 80% of engineering effort on its most automated mechanism.
Why it matters: These agents refine the serving system rather than directly controlling every ad request. Separating online allocation from proposals, experiments, and human approval clarifies where automation operates.
Who should care: Infrastructure teams optimizing inference capacity, recommendation costs, heterogeneous hardware, or engineering experiments.
Brief: Experiment Transcripts Reveal Decision Provenance
What happened: A synchrotron X-ray study records human–agent and agent–agent exchanges to trace decisions across three nanoparticle experiments. The authors report that 59–68% of corrections involved multiple participants as human-defined goals evolved.
Why it matters: Collaboration and correction paths become inspectable research evidence.
Who should care: Researchers studying autonomous labs and multi-agent teamwork.
Brief: Train Retrieval Models for Vector Pooling
What happened: A Sentence Transformers tutorial incorporates pooling into the training objective. In the author’s experiment, a fourfold reduction in document vectors retained 97.8% of the pooling-aware model’s unpooled NanoBEIR nDCG@10, versus 94.6% of the conventionally trained model’s own unpooled score.
Why it matters: Compression can shape training rather than remain a post-training storage step.
Who should care: Engineers optimizing multi-vector retrieval indexes.
Sources
- Gemini Decision Separates Classification From Generation Primary source
- Gemini Decision Multimodal Decision Model — supporting primary source
- Gemini Decision Multimodal Decision Model — supporting primary source
- Qwen4 Preview Is Limited to Token Plan Subscribers Primary source
- AlphaGenome 2 Conditions Variant Predictions on Cell State Primary source
- AgentCore Exposes Web Search as a Standalone API Primary source
- Meta Agents Refine Compute Allocation Primary source
- Brief: Experiment Transcripts Reveal Decision Provenance Primary source
- Brief: Run Time-Series Forecasts Inside Postgres Primary source
- tsfm-pg — supporting primary source
- Brief: Train Retrieval Models for Vector Pooling Primary source
- Sentence Transformers — supporting primary source
- Sentence Transformers — supporting primary source
- Tabletop Turns Group Plans into Shared Interactive Tools Primary source
- Brief: ElevenAI Connects Dictation with Meeting Follow-through Primary source
- Hash LMG Lets Small Models Storyboard While the App Builds the Film Primary source
- Open Hash LMG — supporting primary source
- Brief: Zumungo Turns Home Automation Requests into Reviewable Rules Primary source
- Zumungo — supporting primary source
