Agentic Video Understanding in Gemini

Start here ↓

Google just made long-video understanding agentic.

Gemini’s new agentic video understanding lets the Gemini model decide what parts of a video to inspect through frames, audio, or transcripts. That is very different from sampling a video at a fixed FPS and hoping the important moment made it into context.

It means the model can skim, search, zoom in, rewatch, and pull evidence only when the question needs it.

This feature cuts token usage by up to 88%, lowers cost by up to 66%, and improves accuracy by up to 7%. Now available through the Gemini API in AI Studio and Gemini Enterprise Agent Platform.

Introducing agentic video understanding with Gemini

🚀 Shipped

Anthropic launched Claude Fable 5.1 and Mythos 5.1. Fable 5.1 is the generally available model for coding and knowledge work, while Mythos 5.1 is the restricted version for trusted cybersecurity and life-sciences programs. Anthropic is also cutting cache-read pricing by 75%, which matters for long agent runs that keep reusing the same repo, docs, and tool context.

Cursor already added Fable 5.1 and says it is their strongest model on CursorBench 3.2, scoring 73.4% at max effort.
Claude Fable 5.1 and Mythos 5.1 | Cursor announcement

Dr. Fei-Fei Li’s World Labs introduced Atlas, a world model that cares about camera control. Atlas works across text, images, video, and 3D, and can generate video with precise camera paths instead of vague “pan left” prompting. It can also reconstruct scenes from sparse images or video into point clouds and Gaussian splats, which is the part to watch for robotics, VFX, and game tooling.
Atlas by World Labs

Perplexity’s Hybrid Compute is a great application of “local when it matters, cloud when it helps.” Hybrid Compute lets the app split a task between cloud models and a local model on your Mac. Web research can go to the cloud, but private files, app steps, and PII-heavy work can stay on-device, with an open-sourced classifier deciding what leaves the machine.
Perplexity Hybrid Compute

Shopify open-sourced Tangle for building ML pipelines visually. It gives teams a drag-and-drop editor for experiments, reusable pipeline components, collaborative runs, and caching so repeated steps don’t waste compute. The nice part: it makes the loop inspectable, so agents are not just “trying stuff” in a chat window.
Tangle

Alibaba refreshed Qwen3.8-Max for bigger coding and agent runs. qwen3.8-max-0902 is a 2.4T-parameter model with a 1M-token context window, stronger coding/cowork post-training, and better tool orchestration for long-horizon projects. It is live on QwenCloud at $2/M input and $6/M output tokens, with cheaper cache reads.
Qwen3.8-Max-0902

GojiberryAI open-sourced a Sales OS for Grok Bot. It turns Grok into a 13-agent outbound team for finding intent signals, checking ICP fit, researching accounts, drafting LinkedIn messages, handling replies, and qualifying meetings. It can run autonomously, but starts in “show me the list before anyone is contacted” mode, which is exactly where sales agents should start.
GojiberryAI Sales OS

GitHub CLI can now attach screenshots and videos without opening GitHub. Their latest release adds a repeatable --attach flag for issues, PRs, and comments. This is a small CLI change that fixes a very real agent annoyance: stop describing the broken UI, attach the screenshot.
GitHub CLI media in issues, pull requests, and comments

Reducto shipped r-1, a cheaper parser for ugly documents. The new model is built for the stuff that breaks other OCRs like dense tables, strikethroughs, watermarks, citations, and weird visual layouts. Reducto says r-1 preview cuts error rate 20% versus its legacy agentic OCR models and costs 1 cent per page.
Introducing r-1

Hermes Agent v0.21.0 is basically agent society plumbing. The Pantheon release adds bot-to-bot DMs across profiles and gateways, cron jobs with continuity and monitor-mode suppression, live steering for delegated subagents, JSON schema validation, and a much better MCP command center. If you are running multiple agents already, this is the release where handoffs start feeling less like duct tape.
Hermes Agent v0.21.0

OpenClaw 2.0 is incredibly easy to use. Cleaner setup that can reuse your existing ChatGPT/Claude/API/local model access, and a rebuilt browser app that opens straight into a working agent workspace. Other interesting things include shared cloud sessions, so someone else can join or take over live agent work without losing the context. A lot of other QoL updates!
OpenClaw 2.0, Accidentally

🧠 Worth Knowing

This is one of the best agentic engineering setups you’ll read today for working with multiple coding agents. David Ondrej is running agents across bb, cmux, Ghostty, Herdr, VPS boxes, and subscription plans, then prioritizing finished work with Corral. You probably do not need the whole stack, but if your current workflow is “check five agent windows randomly,” this is worth skimming.
David Ondrej’s agentic engineering setup

Poteto published the workflow stack she uses to keep Grok Bot coding runs from turning into slop. Lauren Tan works on Grok Bot, and her pstack plugin for Grok Bot gives your agents reusable skills for verification, feature maps, playbooks, and multi-model review. If you are trying to run more coding agents without drowning in sloppy diffs, this is worth checking.
pstack for Grok Bot | pstack repo

A new paper takes aim at the laziest safety phrase: “human in the loop.” Researchers argue that human oversight is not a checkbox, especially when agents hide intermediate steps, produce too much to review, and slowly train users to stop checking.
AI Agents Push Humans Out of the Loop

Simon Willison wrote the missing manual for ChatGPT Work. According to him “it's a deeply confusing but extremely powerful tool with a whole lot of useful features that aren't available in regular ChatGPT”
Understanding ChatGPT Work

TinyFish gave DeepSeek Harness free Search and Fetch. Install the TinyFish CLI and DSH picks it up automatically, which means agents can search the web and fetch clean page text for free when stale context is not enough.
TinyFish announcement | TinyFish CLI

🔧 Clone and Run

Clone & Run of the Day
Diffusion Studio is an open-source video editor built for agents. The simplest way to think about it: an IDE, but it renders a video canvas instead of text. Every edit is code, so an agent can cut, tweak, and reuse video edits without handing you one frozen MP4.
Diffusion Studio Editor

Obscura is a headless browser for agents. It gives agents a browser they can drive through CLI or MCP, with tools for clicks, forms, screenshots, PDFs, JavaScript, network logs, and scraping. Worth trying if your agent keeps getting stuck on websites that plain fetch cannot handle.
Obscura

Rakazo is an open-source Grok Bot-style app. You get persistent AI teammates with their own memory, routines, browser, terminal, files, and computer access. You can bring your own models and run the stack yourself instead of depending on one hosted agent app.
Rakazo

ECC is a giant repo of agent skills, configs, and workflows. It has setup files for Claude Code, Codex, OpenCode, Cursor, Gemini, Hermes, Qwen, and more. Would treat it as an awesome reference library first, not something to drop blindly into a real repo.
ECC

headcount turns Claude Code into an agent org chart. Instead of one giant coding agent, it gives you departments like engineering, security, finance, legal, growth, and support, each with its own skills.
headcount

VoiceStudio is a local-first ElevenLabs alternative. It runs voice cloning, dubbing, dictation, transcription, and audiobook workflows across Mac, Windows, and Linux. No account or API key needed for the core workflow
VoiceStudio

Awesome LLM Apps is a curated collection of 100+ AI Agents, Agent skills, and RAG apps. It covers models from OpenAI, Anthropic, Google, and open-source models like GLM, DeepSeek, and Qwen that you can run locally on your computer. (Now accepting GitHub sponsorships)

📊 By the Number

Number of the Day
Together AI cut dedicated H100 inference from $5.49/hr to $3.99/hr for September. The discount applies automatically to new and existingdedicated inference deployments, so teams already running H100 endpoints get the lower bill without moving anything.
Together AI announcement

A 2B model was trained from Qwen2-1.5B on a single RTX 5090 inside a $5,090 budget. Puro-2B is an open recipe for small-model pretraining on consumer GPUs, with data, code, and weights released under Apache 2.0. The caveat is that the budget and Qwen2-1.5B comparison come from the paper’s own evaluation setup.
Puro-2B paper

Someone benchmarked Qwen3.8-Flash-Next across the whole local hardware ladder. In llama.cpp, the model goes from 8.34 tok/s on CPU-only to 109.07 tok/s on a 96GB VRAM setup. The ceiling comes from an RTX 6000 PRO-class card, so this is a hardware map, not a casual laptop win.
Qwen3.8-Flash-Next llama.cpp thread

That's all for today. Come back tomorrow for the next batch of AI tools, model drops, agent repos, and weird benchmarks worth your time.

If you found one thing to try, share the issue with someone who ships.

Unwind AI - X | LinkedIn | Threads

Awesome LLM Apps | Sponsor Us

Cut Lead Review From Hours To Minutes

Sign up for a free trial of Attio, the agentic CRM.

Ask Attio to build a daily workflow that surfaces the deals that need your attention today, like anything with a stage change, a recent reply, or a new signal in the last 24 hours.

Review your pipeline in Claude, synced live from Attio via MCP.

That's it.

Try Attio Now