Explore what's new in AI

Study reports reviewer agents add 5.9–9.5 Terminal-Bench points

A recap of an Alibaba–Cornell study reports reviewer agents add 5.9–9.5 Terminal-Bench points at 1.13–1.25× cost. The study compares 17 verification designs across seven models.

Paper plans React components that agents can read and edit directly

Paper's upcoming component workflow aims to keep code and canvas designs in sync, with agents editing components as React. It could reduce friction between design iteration and implementation.

OpenCode team starts a project for cross-platform native apps

The team is using AI to build native apps and refine platform-specific details such as macOS font rendering. The project offers a practical example of agents contributing directly to shared product development.

CopilotKit is building a Flutter SDK for agent-native mobile apps

The planned SDK uses the AG-UI Dart protocol to bring in-app copilots, native generative UI, and human review flows to iOS and Android from a shared codebase.

Together AI maps a composable open-source AI stack

Separating models, inference, routers, harnesses, and tools lets teams swap components independently, making controlled cost and latency comparisons possible without rebuilding workflows.

Claude Code turns long documents into short Remotion video recaps

A reusable prompt converts dense plans, transcripts, or research into a narrated, shareable video, helping teams review long outputs without reading every section.

Kubernetes control loops inspire an agent workflow for keeping forks current

A proposed design-iterate loop combines upstream drift concepts with agent skills to preserve local changes while regularly updating a fork, reducing manual merge work.

RepoPrompt CE 1.7 adds granular permissions and session recovery

The update improves multi-model workflows with clearer Oracle results, MCP image support, draft refinement, shortcuts, and recovery tools for interrupted sessions.

Yuchen Jiang routes AI-written code through an AI review agent

With AI writing 95% of the team’s code, an agent reviewer addresses approval bottlenecks and offers a concrete pattern for scaling code review.

BAMA bundles 21 agent skills for rebuilding Mac apps in SwiftUI

The workflow moves from screenshots and user stories to one-story-per-commit implementation, tests, and light/dark snapshots, making app reconstruction more systematic.

Entire lets developers claim GitHub usernames for forge migration

A portable handle could make moves between code-hosting platforms less disruptive by preserving a developer’s identity across forges.

Amp releases an Android app for following coding-agent work

The mobile client brings agent conversations, changes, and project artifacts to Android, extending coding-agent workflows beyond a desktop terminal or browser.

A compute-time perfect hash can speed up known-string lookups

A practitioner highlights a technique that computes a perfect hash for a fixed set of strings, enabling fast lookups and string equality checks without paying runtime hashing costs.

Tigris explains how it builds transactional indexes on FoundationDB

FoundationDB does not provide relational-style indexes out of the box. Tigris’s design walkthrough covers transactional index maintenance and the database’s five-second and 10 MB transaction limits.

T3 Code 0.1.3 cuts thread-loading and connection data transfer

The update says opening a thread loads 85% less data and cold connections send up to 93% less, alongside faster search, lower agent-instruction overhead, and broader file previews.

Zed adds shared Claude Code sessions through DeltaDB's /delta share command

DeltaDB's /delta share command lets teammates join a Claude Code session with its context and uncommitted work. Zed also released a Delta CLI for tool integrations, with Zed and Pi support planned.