Explore what's new in AI

OpenRouter's playground sidebar adapts to the model type

The new sidebar makes the playground easier to use across different model families, so selection and prompt setup take fewer clicks. Small UI tweak, real workflow win.

Claude renames extra usage to usage credits

The rename reflects a broader model where credits now buy premium features like Fast mode, not just overage. That changes how teams think about spend and limits.

Google Search rolls out Preferred Sources globally with 345k sources selected

Readers can now pin their preferred sources and reportedly click through more often after doing so. That turns source curation into a product feature, not a browser hack.

OpenAI's Codex desktop app keeps running on a Mac from mobile

Remote connection plus keep-awake mode lets the Mac do the heavy lifting while you work from your phone. That's a nice mobile handoff for long-running agent sessions.

Claude Code makes Fast mode the default for Opus 4.7

Fast mode trades a higher per-token rate for roughly 2.5x speed, which is exactly what you want for live debugging and rapid iteration.

Claude Design doubles token limits across every plan

Anthropic is pushing higher limits into Claude Design, which should make longer sessions and bigger artifacts less annoying. Useful for teams hitting ceilings on real work.

Codex adds a popout hotkey and projectless chat for quick questions

A global shortcut plus projectless mode makes Codex feel like a fast scratchpad when you need an answer without repo context.

OpenRouter shows how to build long-horizon agents with resumable loops

The example wires ceilings, persisted state, and progress streaming together, a solid pattern for agents that need to work for hours.

Alpha-Eval turns agent evaluation into a multi-player training game

The framework uses proposer, judge, agent, and builder roles to generate training signal, but it still has reward-hacking and collapse risks.

Hermes Agent v0.14.0 adds OAuth, local proxies, and Windows support

The release widens Hermes across more apps and platforms, while tightening browser and file workflows. OAuth, local proxies, and Windows support make it much more usable.

MaxSim ships a Hugging Face kernel for late-interaction retrieval

MaxSim attacks the bottleneck in late-interaction retrieval, where materializing similarity scores gets expensive. That can make reranking and hybrid search much cheaper.

Browse.sh open-sources agent skills for repeatable browser tasks

Browse.sh turns repeatable browser tasks into reusable skills and feeds the results back into a knowledge base. That's a practical pattern for taming messy web automation.

Prime Intellect shows how to run agents on fully synthetic RL environments

Automating the RL environment itself is a big step toward automating RL research. Fully synthetic environments could make agent training scale without waiting on human labels.

AI Toolkit adds LoRA merging and scriptable frontend actions

Merging recent checkpoints can behave like a poor man's EMA, but the bigger win is scriptable frontend actions. That opens up more in-app automation for model tooling.

Tejas Kumar's browser agent demo shows harnesses beat prompt tweaks

A cheap browser agent only works once the harness caps loops, verifies tool calls, and handles login state. The lesson is that guardrails can beat prompt tuning.

Agentic Systems as Boosting Weak Reasoning Models reaches frontier SWE-bench with selection

The paper shows a weak model's top-8 proposals can match frontier SWE-bench if you select and verify well. That shifts value from raw model size to orchestration quality.

Gemini CLI scores 43 on the Coding Agent Index, far behind the leaders

A 43 on the Coding Agent Index puts Gemini CLI well behind the leaders. The gap matters because coding agents are judged on workflow, not just raw model quality.

Meta's AIRA splits architecture search into planning and implementation agents

Splitting architecture search into planning and implementation agents beat a single end-to-end agent on a real task. That's a strong template for other agent pipelines.

Claude Code's large-codebase guide maps CLAUDE.md, skills, hooks, MCP, and subagents

The table clarifies where project memory, reusable expertise, automation, and tool access should live. That helps teams keep large codebases navigable as agents get smarter.

Hugging Face's RL for Agents workshop covers the open-source post-training gap

The workshop explains where RL-for-agents is still brittle and what open-source stacks are missing. Good background before you try to post-train your own agent loop.

Black Forest Labs ships FLUX MCP for image generation inside agents

The MCP wrapper lets agents route image requests to the right FLUX model, from quick drafts to production assets. That makes creative workflows much easier to standardize.

Gabriberton argues Python code for agents should use assert statements

Asserting expected types early gives coding agents a fast, local failure instead of a weird downstream bug. It saves time, tokens, and a lot of debugging ambiguity.

React Review Bot flags bad hooks and race conditions before merge

Static review gets an agentic pass that can catch hook misuse, UI race conditions, and perf regressions before merge. Useful if human reviewers are missing the same bugs repeatedly.

Neo4j's agent-memory repo models short, long, and reasoning memory as a graph

The repo treats memory as a graph problem, not just RAG. That lets agent state span short-term chat, long-term entities, and reasoning traces without flattening everything.

Lovable lets skills auto-apply and be called from slash commands

Skills now auto-apply when relevant and can be invoked directly, which makes reusable agent behavior easier to package. Good step toward less prompt copy-paste.

Zed adds Git history and custom git commands to the editor

The editor is moving more git flow into the UI, with history inspection and custom commands. That cuts context switching for teams living in the repo all day.

Hermes Agent Kanban automates prompt triage into subtasks and profile assignment

One prompt can now fan out into subtasks and auto-assign profiles, so triage work turns into coordinated execution. Profile descriptions should make routing less brittle.