One-Line Summary

Google shifts AI from chat to action with Gemini 3.5 Flash, while open-source tools help feed agents cleaner data and give teams blueprints to deploy them.

LLM & SOTA Models

Google launches Gemini 3.5 Flash for faster agent work

Google’s new Gemini 3.5 family is built to execute complex, multi‑step tasks and coding workflows — not just answer questions — and the first model, 3.5 Flash, is available broadly: it’s the default in the Gemini app and AI Mode in Search, accessible to developers via Google Antigravity and the Gemini API in Google AI Studio and Android Studio, and to enterprises through the Gemini Enterprise Agent Platform and Gemini Enterprise. ¹

On numbers, 3.5 Flash outperforms Gemini 3.1 Pro on Terminal‑Bench 2.1 (76.2%), GDPval‑AA (1656 Elo) and MCP Atlas (83.6%), leads in multimodal understanding (84.2% on CharXiv Reasoning), and generates output tokens about 4× faster than other frontier models — often at less than half the cost, according to Google. ¹

That speed‑performance balance targets “long‑horizon” agentic work: paired with the Antigravity harness, organizations deploy collaborative subagents to rename and categorize assets, refactor legacy codebases to Next.js, or even synthesize a research paper into a playable game in six hours. Partners like Shopify, Macquarie Bank, Salesforce, Ramp, Xero, and Databricks are applying these workflows to forecasting, onboarding, enterprise task automation, invoice OCR, tax prep, and data operations. ¹

Google also introduces Gemini Spark, a personal AI agent powered by 3.5 Flash that runs 24/7 under user direction across Workspace and third‑party apps; it’s rolling out to trusted testers and Google says it plans a beta for Google AI Ultra subscribers in the US, alongside strengthened safeguards under its Frontier Safety Framework to reduce harmful content and mistaken refusals. ¹

Open Source & Repos

Firecrawl adds file parsing for agent‑ready documents

Firecrawl is a toolkit that lets AI agents search, scrape, and clean the web; the latest v2.10 adds a /parse endpoint to upload local files up to 50 MB — including PDF, DOCX, ODT, RTF, XLSX, HTML — and get back clean Markdown, JSON, or a summary with preserved tables and reading order. ²

For teams wiring agent pipelines, this replaces brittle one‑off parsers with an API that returns predictable, clean text; enterprise plans include Zero Data Retention to keep uploads from being stored. ²

Pi agent harness bundles a coding agent and unified model API

Pi is an AI agent toolkit that ships an interactive coding agent command‑line interface (CLI), a unified large language model (LLM) application programming interface (API), text user interface (TUI) and web UI libraries, a Slack bot, and deployment pieces. ³

The v0.75.3 update fixes HTTP/2 session crashes in the Node CLI by reverting to the previous HTTP/1.1‑only fetch dispatcher — a stability tweak that matters when your agent runs long, multi‑step jobs. ³

Netron updates model viewer for ONNX, PyTorch, and more

Netron is a visual viewer for neural network and machine learning models — you can open files in the browser and inspect layers, tensors, and shapes across formats like Open Neural Network Exchange (ONNX), TensorFlow Lite, PyTorch, Core ML, OpenVINO, Keras, Caffe, and more. ⁴

Release v9.0.8 refreshes the cross‑platform apps, and the project also offers a one‑click browser version for quick inspections when you don’t want to install anything. ⁴

Nvidia’s video search blueprint shows how to build vision agents

Nvidia’s Video Search and Summarization blueprint is a suite of reference architectures for GPU‑accelerated computer vision agents — combining accelerated vision microservices, vision‑language models (VLMs), and large language models (LLMs) to search and summarize content from video. ⁵

Aimed at developers building production apps, the repo lays out agent workflows, components, and hardware requirements with documentation and a quickstart, giving teams a tested pattern to follow rather than assembling from scratch. ⁵

Community Pulse

Hacker News (512↑) — Speed impresses, but reliability and throttling raise concerns ⁶

"We've been daily-driving this model for a few weeks and let me tell you, everything it does is a lot. Fast as fuck and it's actually not bad intelligence-wise for a fast model. It basically tries to make up for any intelligence deficit by just doing a lot, checking a lot, retrying a lot. That's not to say I don't spend my days raging at it... a lot... but it's not that bad. It does tend to ignore completion criteria but it doesn't obviously degrade when being nudged like some models do." — Hacker News ⁶

"I've had the $20 Gemini plan to use when my local setup runs into tougher problems and the throttling today has been bonkers. I canceled my subscription and will look into upgrading my local setup." — Hacker News ⁶

Why It Matters

AI is moving from chat to action. Google positions Gemini 3.5 Flash as the engine for agents that plan, call tools, and finish tasks, while strengthening safeguards to keep outputs safer in consumer and enterprise settings. ¹

On the ground, reliable inputs and reference designs matter: scraping and parsing clean data (Firecrawl), inspecting models quickly (Netron), and following proven blueprints (Nvidia VSS) help teams ship agent workflows — as users continue to watch for consistent completion behavior and service throttling. ²

This Week to Try

1. Gemini 3.5 Flash as default: open the Gemini app or AI Mode in Search and try a multi‑step task (e.g., draft, revise, and send a doc).
2. Firecrawl /parse: upload a PDF to get clean Markdown/JSON for your agent pipeline — see repo examples at https://github.com/firecrawl/firecrawl.
3. Netron in browser: visit https://netron.app and drop in a .onnx or .tflite model to inspect layers and tensors.