TLDR AI 2026-01-28

Agent A finished, Agent B is waiting, Agent C failed halfway through, and you have no idea what state your system is in. Building agents truly is just the tip of the iceberg.

In enterprise environments, your LangChain agent is almost never the whole story. This Orkes technical guide shows how to coordinate your agents, services and systems in durable, observable workflows.

👉 Instead of stitching everything together with custom glue code, you get production fundamentals built in: stateful execution, retries, parallel steps, human-in-the-loop checkpoints, and full visibility into every run.

The walkthrough includes a real example with four cooperating agents, along with a public GitHub repo you can clone and run yourself.

Read the guide

OpenAI's Prism is a free collaborative scientific writing platform powered by GPT-5.2. It integrates LaTeX-native writing, revision, and publishing workflows in one workspace, aiming to streamline and accelerate research.

Google Search now features Gemini 3 as the default AI model for enhanced AI Overviews, enabling users to find detailed answers quickly. Users can seamlessly engage in follow-up conversations with AI Mode directly from these overviews for more in-depth inquiries. These upgrades aim to provide a fluid search experience, combining quick snapshots with the option for deeper exploration.

Kimi K2.5 is a multimodal model capable of visual reasoning, code generation from UI and video inputs, and agentic task orchestration using a swarm-based architecture. It's built on 15T vision-language tokens through continual pretraining.

An INT4 quantization-aware training pipeline that enables rollout of ~1TB-scale models on a single H200 GPU. By combining fake quantization during training with real quantization at inference, it achieves BF16-level stability and opens a path to cost-efficient, high-performance deployment.

Context graphs aren't retrieval systems; they're production world models. They allow you to ask "what breaks if I merge this PR?" so you can fix defects before they hit production. PlayerZero builds these production world models so your AI actually understands how systems behave.

Explore context graphs →

Learn about PlayerZero →

AI2 released open-source coding agents and a training recipe that makes it affordable and practical to build coding agents for private or internal codebases. Its method reproduces top-tier performance at drastically lower compute costs compared to previous approaches.

SOAR is a self-improvement framework that uses meta-RL to help pretrained models learn from problems they initially fail at. By grounding synthetic problem generation in student model progress, it enables learning on sparse-reward tasks without requiring correct solutions upfront.

The Kimi Agent SDK is a set of multi-language libraries that expose the Kimi Code agent runtime in applications. It can be used to build products, automations, and custom tooling. The SDK clients reuse the same Kimi CLI configuration, tools, and MCP servers to stream responses in real-time, surface approvals and tool calls, and let users orchestrate sessions programmatically. Go, Node.js, and Python are supported.

DeepSeek-OCR 2 is a significant upgrade to DeepSeek-OCR. It maintains high visual token compression while achieving meaningful performance improvements. The model is powered by DeepEncoder V2, which implicitly distills causal understanding of the visual world through the integration of both bidirectional and causal attention mechanisms. This leads to causal reasoning capabilities in the vision encoder and marked lifts in visual reading logic.

Google has introduced Agentic Vision in Gemini 3 Flash. The new capability enables the model to effectively use code and reasoning to improve performance for common vision tasks. It is available to users through the Gemini API in Google AI Studio, Vertex AI, and within the Gemini app. Google plans to extend Agentic Vision's reach by supporting more model sizes and integrating additional tools like web and reverse image search.

SoftBank is in talks to invest up to $30 billion more into OpenAI. OpenAI is seeking up to $100 billion in new capital in a round that could value it as much as $830 billion. It needs substantial funding to continue developing its AI models, pay for its vast computing needs, and retain top researchers in an increasingly competitive market. The company is aiming to raise capital from Middle Eastern sovereign-wealth funds and other venture capital funds.

Stop predicting workflows. Start observing them. BAAPs discover, assemble, and deploy agents automatically - based on how your organization actually works. Learn how.

Clawdbot's persistent memory system, which maintains 24/7 context retention, remembers conversations and builds upon previous interactions indefinitely.

Manus AI's Agent Skills open standard enables modular AI workflows for professionals, integrating Python and Bash scripts with a focus on efficiency and interoperability.

Google has expanded availability of AI Plus, which includes access to Gemini models and AI productivity tools, to 35 more countries.

FLORA unifies the best creative AI models across text, image, and video into one workflow to accelerate the creative process.