World Wire

JetBrains Research Open-Sources KotlinLLM. KotlinLLM is an IntelliJ IDEA plugin for Kotlin/JVM projects that adds a language feature called Smart macros . A Smart macro is a regular Kotlin function call whose body is generated Kotlin code. The public API is deliberately small. asLlm<F, T>(from, hint) converts an input of type F into a typed value T, such as a data class, enum, list, or primitive. mockLlm<T>() generates a stateful implementation of an interface T, whose behavior depends on which methods are called on it. Copy Code Copied Use a different Browser val issuesApiUrl: String = asLlm(repoInput, hint = "GitHub API URL: get all issues, including closed") val issues: List<Issue> = asLlm(response, hint = "Return all beginner-friendly issues for this repository") The runtime loop When a project launches through the KotlinLLM run configuration, the plugin scans for asLlm and mockLlm calls, updates generated bootstrap/provider/...

Nous Research has released Hermes Agent support for Buzz , Block’s open source, self-hostable workspace where humans and AI agents share the same channels. Buzz is built on Nostr. Every message is a signed event on a relay you own, and every participant, human or agent, is a keypair. That removes the bot-token model. Agents get their own identity, their own channel memberships, and their own audit trail. Is it deployable, and for whom? Yes, and both sides are self-hostable today. Buzz is Apache-2.0 with 18.8k stars; Hermes Agent is MIT licensed. Solo developers and small engineering teams can run it through Buzz Desktop with no configuration. Mid-market platform teams are the strongest fit, because the relay sits on Postgres, Redis, and S3/MinIO. Enterprises should scope this as a pilot, since mobile clients and workflow approval gates are still being wired up. Practical applications include incident memory over channel history, branch-as-room code review, agent-drafted...

In this tutorial, we build and execute a multi-agent workflow with Omnigent using a reliable, isolated Python environment created with uv. We configure a financial research lead agent that retrieves a live USD-to-EUR exchange rate from an external API, prepares a concise client-ready summary, and delegates its draft to a dedicated text-auditing sub-agent for clarity and length validation. We define reusable Python functions as callable agent tools, describe the complete agent structure in YAML, and use the Claude Agent SDK as the execution harness. We also manage the Anthropic API key securely through environment variables, apply non-interactive policies that limit tool calls and control session costs, and run the workflow directly from Colab without requiring Node.js, tmux, or an interactive terminal. Through this implementation, we explore how Omnigent combines agents, tools, delegation, live data access, and governance within a single configurable system. Copy Code Copied Use...

Google DeepMind has released Gemini Robotics 2 , the intelligence layer for its next generation of robots. The release moves the stack past table-top manipulation into whole body control, five finger dexterity and multi robot teamwork. It ships as three separate models with three different access tiers. Most robots today are pre-programmed or tele-operated for narrow, repetitive task sequences. They do not adapt to unpredictable environments, and skills rarely transfer between robot bodies. Gemini Robotics 2 targets all three limits at once. TL;DR Three models ship together: a VLA, an embodied reasoning VLM, and an on-device VLA. One checkpoint drives Apollo 2 with two different hands plus a Franka Duo gripper. Gemini Robotics ER 2 is public preview; the VLA and on-device models stay gated. Multi-finger dexterity is still the weak axis, ranging from 32% to 92%. ASIMOV-Agentic, a new safety benchmark, is on Hugging Face under CC-BY-4.0. 3 Models and Wh...

Tencent has released AngelSpec , an open-source, torch-native training framework for speculative-decoding draft models. The release covers both autoregressive multi-token prediction (MTP) and the block-parallel DFlash family. Most speculative-decoding work searches for one drafter that scores well on an averaged benchmark mixture. Real serving traffic does not look like that mixture. AngelSpec treats workload heterogeneity as a first-class design constraint , and specializes structure, training data, and verification depth around it. Why one universal drafter underperforms Speculative decoding is lossless. A lightweight drafter proposes several future tokens, and the target model verifies them together in one forward pass using rejection sampling. Acceleration then depends on two things: how many draft tokens get accepted, and how long the complete draft–verify round takes. Those two quantities move in opposite directions across domains. In high-entropy open-ended con...