Databricks just ran a real-world coding test (not some benchmark

2026-07-21 04:34

Databricks just ran a real-world coding test (not some benchmark suite) pitting different agent tools against various models. Chart’s simple: higher on the Y = better success rate, further left on the X = cheaper. Pi with top models is sitting at 80%+ success while burning way fewer tokens than everything else. Been running Pi myself for a bit now — grok-4.5 + DeepSeek-v4 plus my own setup (main agent orchestrating, subagents running concurrent with isolated context). Barely any rework, stupid fast, finally stopped refreshing my phone waiting for results. Main agent + concurrent subagents is clearly the direction. Seeing more tools start shipping this workflow lately.

Gemini 3.6 Flash is here

. 1M context, 65K output, takes image/audio/video. Google says it uses fewer tokens than 3.5 Flash, writes better code, and reasons better across modalities. 3.5 Pro is still MIA. My bet: they skip 3.x and drop 4.0 Pro to chase the top models.

2026-07-22

Just found a legit 0-cost local solution for Agent search &

crawling I’ve been playing around with this open-source project called wigolo lately. It lets you do full search, scraping, crawling, extraction, and research—all running completely locally. No API keys, no cloud services, nothing leaves your device. This means your agents can search the web and grab pages for free, without relying on Firecrawl, Exa, Tavily, Brave, Jina, or any of those other paid tools. Super clean for regular users too—no more messing around with API keys or environment variables. github: KnockOutEZ/wigolo

2026-07-20

Meituan shipped LongCat-2.5

. 1.6T params, ~48B active, native multimodal, aimed at long-horizon agents, trained start to finish on Chinese hardware. Images get quantized into multi-level discrete tokens, and that's where Visual Coding and 3D pick up. On the RL side, failed trajectories get turned back into verifiable training tasks. W4A4 inference, 43% faster prefill. Meituan's own evals put it between Claude Opus 4.8 and Opus 5.

2026-11-02

SenseTime open-sourced SenseNova Loop-9B-Preview.

It stacks layers instead of parameters. The decoder has 32 layers, but layers 4 through 27 run twice with shared weights, so a forward pass does 56 layers of work while the checkpoint stays at 9B. Their own eval numbers: 71.38 on BIRD-Dev and 13.72 on AutomationBench, both above Qwen3.5-27B. Terminal Bench is still behind.

2026-11-02