agent

This paper introduces TRACE, a dense credit-assignment method for reinforcement learning of long-horizon agents that significantly improves Qwen3 small MoE models on agentic benchmarks without additional critic models.

On the Kimi K3 model, the open-source agent framework Maka achieves a 10% higher overall pass rate on Terminal-Bench 2.1 compared to the official Kimi Code CLI, and 20% higher on hard tasks. Through optimizations like context-budget pruning, streamlined tool surface, and concise system prompts, significant performance gains are realized. The full report and harness are open-sourced.

Kimi K3 demonstrates consistent improvement in end-to-end knowledge work, as evidenced by public benchmarks and internal reviews, indicating enhanced agent capabilities.

A user shares surprise at how well an open-source AI agent works for client research using Hyperagent, which lets you swap between models like Kimi K2.6, GLM 5.2, DeepSeek V4, and Qwen 3.7.

A discussion or tool for visualizing how AI agents use tools, likely to aid understanding and debugging of agent behavior.

OpenAI released a $230 limited-run keyboard, the Codex Micro, co-designed with Work Louder, that serves as a command center for managing AI coding agents in Codex. The device features Agent Keys, Command Keys, a joystick, and a dial for adjusting agent reasoning level.

This post describes how to build a self-improving code review agent as part of a cloud software factory, using a code review skill, GitHub actions, and an outer-loop agent for continuous improvement.

Codex Micro is a physical command center for agentic work, developed by OpenAI and Work Louder, featuring customizable keys, joystick, and RGB status indicators for controlling AI agents.

Coasty is a new API for building computer-use agents, offering task runs, workflows, managed machines, and lower-level prediction primitives.

OpenAI released the latest prompt guide for GPT-5.6 Sol, emphasizing clearly and concisely defining goals, success criteria, and constraints rather than verbose step-by-step instructions, thereby improving efficiency and performance.

AgentKey is a new tool that integrates platforms such as Bilibili, Douyin, Zhihu, Weibo, WeChat Official Accounts, Xiaohongshu, as well as X, Reddit, YouTube, Instagram, etc. It helps AI Agents access real-time internet information, solving the problem of Agents struggling to reach the Chinese internet, and topped Product Hunt's daily leaderboard.

Agent-Reach is an open-source tool that, with a single command, connects AI agents to 14 online platforms (such as Twitter, Bilibili, YouTube, etc.), solving the pain points of web scraping and API payment, allowing agents to directly obtain online information.

Memvid is an open-source tool that packages agent memory into mp4 files, offering significant performance improvements over traditional RAG pipelines and vector databases.

Explores four ways an agent's write can silently disappear, with two detectable and two preventable issues.

A Prime Intellect engineer pointed out that large language models like GPT-5.5 see retrieval accuracy drop from 80% at 256k tokens to 36% at one million tokens, indicating the 'context rot' problem—the model can accommodate but cannot effectively reason over long contexts, posing a challenge to agent applications.

The Cicero project enables self-hosted, bidirectional voice conversation with AI agents, supporting local TTS/STT and integration with various agent protocols and tools like Claude Code and Hermes Agent.

Richard Sutton launches Oak Lab, a new AI research lab focused on real-time learning and planning algorithms, with the ambitious goal of creating a trillion-parameter agent that operates on just 20 watts of energy.

The author argues that giving an AI agent fewer, coarser tools improves reliability and cost by reducing decision surface and preventing the model from confusing factual and simulated outputs.

The author shares their multi-layer memory system built for Agent, including global, project, Founder, and activity memory, and extends it to work logs, planning, queries, and automated report generation.

The author describes building Oogway, a proactive monitoring agent on OpenClaw that investigates anomalies, proposes fixes, and persistently logs each incident into a wiki, creating a compound knowledge base that avoids re-deriving solutions.