AI Briefing — 2026-08-05

🔬 Research

LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks

137 upvotes · Ziyu Ma et al. · Paper

Current LLM agent frameworks accumulate state in a growing context window, causing errors to compound over long multi-step tasks. LongHorizon-Harness externalises task state and introduces a Manage-Execute-Audit (MEA) loop: a manager tracks state, a fresh-context executor handles each subtask in isolation, and a read-only auditor verifies outcomes before the next step. The gains are substantial: Qwen 3.7-Plus jumps from 51.8% → 80.7% on WeaveBench and from 2.8% → 8.3% on OSWorld 2.0; Claude Opus 4.7 goes from 20.0% → 34.3% on an OSWorld 2.0 subset. A lightweight AgentAdapter makes the harness model-agnostic, so these gains transfer without rewriting native agent loops.

DAPD: Dual-Anchored Policy Distillation

73 upvotes · Jianyu Wu et al. · Paper

On-policy self-distillation (OPSD) for post-training LLMs suffers from a privilege illusion: the student mimics teacher behaviour that depends on privileged training-time context it won't have at inference. DAPD fixes this with two anchoring mechanisms — Dual-Path Anchoring (a self-conditioned bridge aligning reference and rollout behaviours) and Dual-Source Anchoring (bidirectional path alignment to reduce reliance on privileged guidance). On Qwen3, DAPD yields +2.69 points at 4B and +2.78 at 32B over standard OPSD on average across tasks — a modest but consistent improvement that matters at scale.

JoyAI-Video-Edit: Real-Time Open-Ended Video Editing with Autoregressive Diffusion

65 upvotes · Yicheng Xiao et al. · Paper · Code

JoyAI-Video-Edit is a 16B-parameter autoregressive diffusion model for causal, streaming video editing — no future frames, no fixed duration. It combines chunk-wise autoregressive generation with Source-Anchored Distribution Matching Distillation (SA-DMD) and Long-Horizon Autoregressive Distillation to prevent temporal drift across long edits. The system achieves 720p editing at ~30 FPS on a single Nvidia B200 GPU, outperforming existing streaming editors while staying competitive with offline systems on both short and long videos.

🛠 Tools

zhaoxuya520/reverse-skill — AI-powered cybersecurity skills router

2,297 stars/day · 18,648 total ⭐ · Repo

A prompt/skill-routing pack for reverse engineering, penetration testing, and security research, designed to plug into AI coding clients (Claude Code, Cursor, Cline, Kiro). It provides AI-powered task routing to the right security subtool, an on-demand toolchain bootstrapper, and a self-evolving knowledge base that persists experience across sessions. The explosive star velocity (~2.3k/day) reflects strong demand at the intersection of AI-assisted coding and offensive security tooling.

lyogavin/airllm — 70B inference on a single 4GB GPU

1,711 stars/day · 28,711 total ⭐ · Repo

AirLLM runs large models on minimal VRAM by streaming model layers from disk rather than loading them all at once — no quantization, distillation, or pruning required. The technique scales impressively: Llama 3.1 405B on 8GB, DeepSeek-V3 (671B) on ~12GB, and Kimi K3 (2.8T MoE) on under 4GB by streaming one expert at a time. The trade-off is speed (not suitable for real-time use), but it makes frontier-scale models accessible for experimentation on consumer hardware.

TencentCloud/TencentDB-Agent-Memory — Team-level shared memory hub for AI agents

1,111 stars/day · Repo

This TypeScript library provides a structured persistent memory layer for multi-agent systems, organising data into four typed assets: Chat Memory, Skill, LLM-Wiki, and Code-Graph. The goal is to make agent knowledge reusable and shareable across agents, frameworks, and sessions — addressing the statelessness problem of most current agent architectures. Being backed by Tencent Cloud suggests production-grade governance and access control are first-class concerns.

📡 Analysis

New release of LLM adds support for reasoning traces, OpenAI Responses, server-side tools, and smarter logging

2026-08-04 · Post · tags: llm, openai, anthropic, llm-tool-use, llm-reasoning, model-context-protocol

LLM 0.32 is a major release of Simon Willison's CLI/Python tool for interacting with LLMs. Key additions: reasoning traces streamed to stderr (keeping stdout clean for piping), native support for the GPT-5.6 family (with GPT-5.6 Luna as the new default), server-side tools like CodeInterpreter and WebSearch via -T, and a new llm openai endpoint command for hitting any OpenAI-compatible API without installation. The redesigned content-addressable SQLite logging and updated Python API (removing the forced conversation abstraction) make this a meaningful upgrade for anyone building LLM-powered pipelines.

llm-anthropic 0.26

2026-08-04 · Post · tags: llm, anthropic, claude, model-context-protocol

The companion Anthropic plugin for LLM 0.32 adds three new Claude 5 models (claude-fable-5, claude-sonnet-5, claude-opus-5) and four server-side tools: WebSearch, WebFetch, CodeExecution, and AnthropicMCP — the last of which allows Claude to call MCP servers server-side in a single API round-trip. Extended thinking is simplified to a thinking_effort option (low/medium/high/xhigh/max); Claude 5 models think by default and -o thinking 0 disables it. Together with LLM 0.32, this makes the CLI a capable, composable interface for agentic Claude workflows.

PipeNetwork/minimax-h3-mlx — MiniMax-H3 text-to-video on Apple Silicon

2026-08-04 · Post · tags: mlx, text-to-video, minimax, generative-ai

MiniMax-H3 is a new omni-modal generative model (text/image/audio/video in → up to 15s video+audio out); this repo ports it to Apple Silicon via MLX. Simon ran it on an M5 Max MacBook Pro: it required ~115 GB of model downloads and took ~45 minutes to generate a single clip, so this is firmly in the "experimentation" category for now. Worth bookmarking as an early signal that large video-generation models are becoming runnable locally, even if slowly.

Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net