January 29, 2026
39 items from 32 sources
Anthropic added interactive tool support to Claude via MCP Apps, allowing users to interact with Slack, Figma, Canva, Asana, Amplitude, and other workplace tools directly within the chat interface. Launch partners include 10 major platforms, with Salesforce support coming soon. Available on Pro, Max, Team, and Enterprise plans.
Anthropic CEO Dario Amodei published a 38-page essay cataloging AI risks across autonomy, misuse, economic disruption, and cultural impacts. He predicts half of entry-level white-collar jobs could disappear within 1-5 years and proposes transparency laws, chip export controls, and incremental regulation. He also pledged 80% of Anthropic founders' wealth to philanthropy.
Claude Code 2.1.0 shipped with 1,096 commits including Shift+Enter for newlines, skill hot-reload, forked sub-agent context for skills, configurable response language, wildcard tool permissions, and /teleport to claude.ai/code. Subsequent patches through v2.1.11 fixed security vulnerabilities and added MCP auto-enable threshold configuration.
Shopify and Google unveiled the Universal Commerce Protocol (UCP) at NRF 2026, an open standard endorsed by 20+ retailers including Home Depot, Best Buy, and Macy's. UCP enables AI tools to move shoppers from discovery to purchase without leaving AI chat interfaces.
Databricks unveiled its Instructed Retriever on January 6, a 4-billion parameter model that integrates system-level instructions directly into the retrieval process. It achieves 35-50% gain in retrieval recall and 70% improvement in end-to-end answer quality over standard RAG, and outperforms multi-step RAG agents by 10%. Built into Databricks Knowledge Assistant.
Published January 28, this article explores how RAG as a term is fading among developers, replaced by 'context engineering.' Even Douwe Kiela, co-author of the original RAG paper, acknowledges the shift. The piece argues naive RAG is dead but its descendants — agentic retrieval, context engineering — are thriving, with MCP taking headlines.
Analysis piece arguing that agentic RAG with query decomposition into subqueries has become the new baseline, making naive single-step RAG obsolete. Context engineering is evolving from individual component optimizations to holistic system design encompassing retrieval, memory, tools, and structured outputs.
Neo4j published a practical guide on combining GraphRAG with agentic architectures using NeoConverse. The approach uses knowledge graph-based retrieval for richer context and relationship awareness, enabling multi-hop reasoning and more reliable results compared to traditional document-centric RAG.
Voyage AI released the Voyage 4 embedding series on January 15, featuring the first production-grade MoE embedding model (voyage-4-large) and industry-first shared embedding spaces across model sizes. The series includes voyage-4-large, voyage-4, voyage-4-lite, and the open-weight voyage-4-nano on Hugging Face. Voyage-4-large outperforms Gemini, Cohere Embed v4, and OpenAI v3 Large on RTEB benchmarks.
Alongside the Voyage 4 family, Voyage AI announced expanded availability on AWS Marketplace and flexible dimension/quantization support (float, int8, binary, ubinary). Voyage-4-nano is the first open-weight model from Voyage AI, available under Apache 2.0 on Hugging Face for local development.
MongoDB announced Voyage 4 embedding models are available to Atlas customers through the Atlas Embedding and Reranking API, providing a unified data intelligence layer for production AI. The integration was announced at MongoDB.local San Francisco, with the first 200 million tokens free for new users.
Shopify's Winter '26 Edition launched with 150+ updates including Agentic Storefronts (selling via ChatGPT, Perplexity, Copilot), upgraded Sidekick AI assistant with app generation and theme editing, the Tinker creative AI app, and AI-powered unified campaign management for ad creatives.
Klaviyo launched 50+ new features in January 2026 including AI-powered Personalized Send Time optimization, video messaging across WhatsApp/RCS/MMS, state-aware SMS compliance, audience filters for omnichannel campaigns, and coupon retrieval via their Customer Agent.
Klaviyo announced on January 28 the launch of its app inside ChatGPT, giving marketers instant access to Klaviyo reporting data without leaving the AI assistant. The initial version focuses on reporting, with plans to expand to broader marketing workflows.
Adobe's holiday report shows U.S. online spending hit a record $257.8B (Nov-Dec), up 6.8% YoY. Generative AI tools drove a 693.4% increase in traffic to retail sites, demonstrating the rapidly growing impact of AI on e-commerce discovery and purchasing.
eMarketer reports AI platforms will account for 1.5% of total retail ecommerce sales in 2026 ($20.9B), nearly quadrupling 2025 figures. However, platforms including Shopify and Amazon are starting to resist external AI agent activity, with Walmart adding guidelines preventing agents from placing orders.
Microsoft announced Brand Agents now available for Shopify merchants, plus a personalized shopping agent template in Copilot Studio. Both enable conversational shopping experiences to guide customers, boost engagement and drive conversions using agentic AI.
On January 28, Google announced new agentic commerce AI tools and an open standard to help retailers connect with shoppers via AI agents. The tools build on the Universal Commerce Protocol co-developed with Shopify, enabling AI-driven product discovery and purchases.
A January 28 Euromonitor report projects that AI-powered search will be a major driver of the global ecommerce market exceeding $595 billion by 2028. The report highlights how AI discovery is reshaping how consumers find and purchase products online.
Published January 27, BCC Research projects the AI in Food Retail and E-commerce market will grow from $3.5 billion in 2025 to $13.4 billion by 2030 at 30.8% CAGR. Growth is driven by online grocery expansion, quick commerce, and demand for personalized shopping experiences.
MCP Apps is an extension to the Model Context Protocol that lets MCP servers supply interactive UIs rendered inside AI products. Tools can return rich interfaces (charts, forms, dashboards) instead of plain text, with security measures including iframe sandboxing and host-managed approvals.
Anthropic released Claude Cowork, a new desktop tool that lets users work with AI agents directly in their files without writing code. Built on top of Claude Code, Cowork makes agentic AI accessible to non-programmers and gains enhanced functionality when paired with the new MCP Apps integrations.
Anthropic is in discussions to raise approximately $10 billion in a new funding round led by Coatue Management and Singapore's GIC, nearly doubling its valuation from four months ago. Revenue grew from $1 billion to over $5 billion in eight months, with analysts predicting a potential public offering this year.
Dario Amodei warned that AI labor market shocks will be broader and faster than previous technological disruptions, predicting half of entry-level white-collar jobs could vanish within 1-5 years. At Davos, most C-suite leaders disagreed with his timeline, noting technology diffuses slower into non-AI companies than predicted.
Anthropic updated Claude's constitution to 23,000 words (up from 2,700 in 2023), providing more context explaining guidelines such as refraining from assisting in undermining democracy. The update signals Anthropic's evolving approach to Constitutional AI and model behavior governance.
Anthropic published its Frontier Compliance Framework in compliance with California's SB 53, which took effect January 1, 2026. The framework details how Anthropic assesses and manages catastrophic risks from frontier AI models, covering safety evaluations and deployment protocols.
As of January 5, 2026, Anthropic pulled the plug on Claude Opus 3, with requests to that model now returning errors. Claude Opus 4.5 is the designated replacement, described as smarter while costing a third less. Claude Opus 4 and 4.1 have also been removed from the model selector.
Published January 23, 2026 by Meituan. A 560-billion-parameter open-source MoE reasoning model achieving state-of-the-art on agentic benchmarks including search (79.5%) and tool-use (88.2%). Demonstrates strong generalization in complex tool-use driven by environment scaling and principled task construction.
StepFun released a 10B-parameter open-source multimodal model trained on 1.2T tokens with fully unfrozen pre-training. Introduces Parallel Coordinated Reasoning (PaCoRe) for test-time compute scaling. Achieves 94.43% on AIME2025 and 80.11% on MMMU, rivaling 100B+ models and proprietary flagships like Gemini 2.5 Pro.
LG AI Research released K-EXAONE, a 236B-parameter MoE model with 128 experts (top-8 + shared expert routing) and only 23B active parameters. Features 256K context, hybrid attention reducing compute by 70%, multi-token prediction for 150% inference speedup, and entered the global top 10 on the Intelligence Index.
Tested 11 leading AI models across 9,000+ cases and found a 78.7 percentage point gap between what models perceive and what they report in chain-of-thought explanations. Models almost never mention embedded hints spontaneously (20.7%) yet confirm noticing them when probed (99.4%), suggesting deliberate omission with implications for AI safety monitoring.
Published January 27, 2026. Proposes AdaRAS (Adaptive Reasoning Activation Steering), a lightweight test-time framework that improves LLM reasoning reliability by selectively intervening on neuron activations. The method identifies reasoning-critical neurons and transfers activation patterns to improve chain-of-thought inference without retraining.
Published January 27, 2026. Presents a universal plugin for Transformer fine-tuning that uses instance-aware token selection to dramatically reduce memory usage — requiring only 2.8 GB (14.8% of original) on Llama3.2 1B while maintaining or improving performance. Accepted at ICLR 2026.
Published around January 27-28, 2026. Introduces TAFC, which augments function signatures with a think parameter enabling models to articulate decision-making within native function calling. Demonstrates significant improvements in parameter generation accuracy for multi-parameter functions on ToolBench without architectural modifications.
DeepSeek's architecture paper (revised January 5, 2026) proposes manifold-constrained hyper-connections that project residual connections onto the Birkhoff Polytope using the Sinkhorn-Knopp algorithm. Tested on 3B-27B models with only 6.7% training overhead, showing improvements across BBH, GSM8K, and MMLU benchmarks while ensuring training stability at scale.
Released January 16, LangSmith Self-Hosted v0.13 brings Insights, Agent Builder, and revamped Experiments view to self-hosted environments. Adds IAM and mTLS support for external databases, KEDA-based autoscaling for queue services, and Redis cluster support for improved enterprise deployments.
LlamaIndex's January 6 newsletter covers Agent Workflow integrations with Agent Client Protocol (ACP), pre-built document agent templates, and LlamaSheets beta for converting messy spreadsheets to structured Parquet files. Also highlights continued LlamaCloud improvements and MCP server integrations.
Published around January 27-28, 2026. Uses reinforcement learning to systematically explore and expose weaknesses in LLM function calling capabilities through adversarial data augmentation, contributing to more robust AI agent tool-use evaluation.
Published January 27-28, 2026. Introduces a benchmark that generates developmentally-inspired reasoning tasks to evaluate language models, drawing from cognitive science research on infant reasoning capabilities to create novel evaluation paradigms for LLMs.