THE OFFICE Agent Harness

Start here ↓

Someone just built THE OFFICE theme agent harness and made it open-source.

Munder Difflin wraps the CLI agents you already use, like Claude Code, Codex, Gemini CLI, Kimi, Grok, OpenCode, gives them memory, wires them into a hive mind, and puts your clone in charge. Michael (ofc had to be him) is the one you talk to to get things done.

It works with subscriptions and keys you already have, keeps the local version on your machine, and lets agents coordinate without everyone pushing into the same shared mess.
github.com/chaitanyagiri/munder-difflin

Also, do not skip By the Numbers today 🤫 

🚀 Shipped

Cursor Cloud Agents can now start without a repo. You can prompt a new web app from scratch, preview it in the browser, then save the code to a Cursor Origin repo when it is worth keeping. If you connect Vercel, Cursor can also publish the app to a live URL.
cursor.com

Vercel made WebGPU easier for agents to test. vgpu lets agents validate shader code in headless Node.js and CI, even when the sandbox does not have a GPU. Useful if you want agents working on visual code without relying on manual screenshot checks.
vgpu.sh

The Claude Code team fixed a quiet cache bug that could cost money in long sessions. The changelog says tool definitions were being re-rendered after OAuth token refreshes, causing a prompt-cache miss roughly once an hour. If you run long agent sessions, that is the kind of invisible leak worth upgrading for.
code.claude.com

Google made Gemini Omni cheaper to iterate with. Gemini Omni 1.1 Flash can now generate rough 360p video drafts before you pay for a higher-resolution render. These drafts are up to 60% faster and cost 1/3rd as much as 720p, so you can test a bunch of directions quickly, then upscale the one you want to keep.
blog.google

Firecrawl added OCR to anydoc for scanned docs. anydoc already turns office files and text-based PDFs into clean Markdown for agents. The new Firecrawl OCR option covers scanned pages too, with sub-5ms handling for docs that do not need OCR and 190ms median per OCR page. It is free to use with no API key.
github.com/firecrawl/anydoc

Agno, the open-source AI agent framework, released its v3.0, shipping with an SDK, runtime, and control plane together. So teams can build, run, and monitor agents in the same stack. Worth a look if you are already comparing agent frameworks for production use.
docs.agno.com

🧠 Worth Knowing

The best RAG stack might still start with plain search. Rafael Pierre’s RAG breakdown is a good antidote to overbuilt retrieval systems: start with BM25, add query rewriting, then move to hybrid or pre-embedding only when the data proves you need it. Remember the 80/20 rule: do not build the 5% solution for a 60% problem.
lighthousenewsletter.com

Anthropic is testing a standard for agents controlling lab hardware. The Model Hardware Standard is a research preview for letting agents operate programmable devices like microscopes, liquid handlers and robotic arms through shared driver primitives. They say that hardware integrations that usually take weeks or months could drop to hours or minutes.
anthropic.com/news

Alibaba’s latest model Qwen3.8-27B looks fine at 4-bit, but falls apart at 1-bit. Quesma benchmarked several GGUF quantizations and found the 17GB Q4_K_M version holds up surprisingly well, while 1-bit collapses to random-chance territory on GPQA Diamond. If you are picking a local coding model for a 24GB card, this is a practical result, not a philosophical debate about quantization.
quesma.com

Someone measured Claude-ish vocabulary across 47,000+ GitHub PRs. A word cluster that didn't exist in 2025 is now 45% of human-authored PRs. And the top word is "load-bearing"
louisabraham.github.io/load-bearing

Anthropic says code review cannot stay line-by-line. Their latest AI-native SDLC playbook argues that once agents write large chunks of code, the bottleneck moves to planning, testing, security, and deployment. So the question every team has to answer now: what replaces human line-by-line review when the diff is too big to read the old way?
claude.com/blog

Terminal-Bench-Science is trying to measure agents on real scientific work. The benchmark uses workflows contributed by working scientists, then grades concrete artifacts like analyses, simulations, proofs, code, and data products. The first release has 70 tasks, and Claude Opus 5 tops the leaderboard at only 30%, which is exactly why this is more useful than another easy eval.
terminal-bench-science.ai/announcement

🔧 Clone and Run

Clone & Run of the Day
Tare analyzes a Claude session and shows where the tokens actually went. If your Claude Code quota disappears in ten minutes, attribution is more useful than another complaint thread.
https://github.com/kelviq/tare

OpenSEO is an open-source alternative to Semrush and Ahrefs. It exposes an MCP server so AI agents like Claude Code, OpenClaw, and Hermes can use your SEO data directly. Agent Skills are reusable workflows that guide your agent through SEO tasks using the MCP.
github.com/every-app/open-seo

Concord is an MCP server for letting Claude Code, Codex and Cursor send messages to each other, live. If your current multi-agent coordination layer is “write to a shared file and hope,” this is worth reading.
github.com/Get-Concord-AI/concord-mcp

Claude quickstarts now include a cookbook for running Claude Managed Agents with Vercel’s Chat SDK and delivering them into Slack, WhatsApp, Discord and Teams. Useful if the thing you keep rebuilding is the chat delivery layer, not the agent itself.
github.com/anthropics/claude-quickstarts

Experiential is an open-source model gateway for agent workflows. It gives you one OpenAI-compatible API across hosted, BYOK and local models, plus controls for who can use which model and how much they can spend.
github.com/experientiallabs/experiential

Awesome LLM Apps (134k+ 🌟 ) 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
Nvidia agrees to buy Hugging Face for $12.9B. No signed agreement yet, but it is big enough to make people nervous because Hugging Face is not just another AI company. It is where a lot of teams host models, datasets, Spaces, and demos. If the platform changed ownership, the question is less “is open source dead?” and more “how much of your workflow depends on one host?”
techcrunch.com

Thinking Machines is giving up to $50K in credits for open-weight safety research. The grants are for projects using Tinker to study things like safer open models, hazardous-data filtering, tamper-resistant safety training and reward hacking. Good fit if you are doing actual experiments on open-weight model safety and need compute credits.
thinkingmachines.ai

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