Can Agents improve their own harness on unseen tasks? The @GoogleResearch team released Harness², a framework that allows improve an harness around a frozen model. How it works: 1️⃣ Agent runs the base harness on each task to record a baseline trajectory 2️⃣ Editor agent inspects the trajectory, drafts 2 candidate edits (SI, Skills, hooks,….), and runs both. 3️⃣ Composer agent diffs the new trajectory against the baseline, merges edits that fixed failures (within a cap), and runs the final harness. 4️⃣ Reflector agent logs verified lessons into file and saves the winning components to a library for future tasks to pull on demand. Results: - 💡 Gemini 3.5 Flash +8.4 on WorkBuddy-Bench and +7.5 on JobBench. - 🛠️ Skills and scripts drive ~50% of edits; dropping skills costs 5.3 pts. - 📚 Gemini 3.5 Flash improves on all 3 benchmarks when editing its own harness. Paper: arxiv.org/abs/2610.14224 Code: github.com/google-research/h…

The new Gemma docs! gemma.ai/ 🥁🥁🥁 From capabilities, to fine-tuning guides, inference optimizations and educational content on how the models work. All at central place!

Gave Gemini 4 Argon a photo of the Manhattan skyline asked it to recreate it as oil painting. It wrote a multi-pass canvas renderer that starts with broad background strokes, blocks in the buildings and river, and finishes with fine specular highlights on the water.

Gemini 4 Argon is now live in the Gemini API and @GoogleAIStudio 🎉 What it has done at Google so far: - A months-long Chrome refactor at ~4x the usual pace - 35%+ more agentic code submissions - Improved session creation in ADK by 40% - 2.9x the throughput increases in SGLang for Gemma 4 With Gemini 4 you should spend most of your time writing a clear goal and a way to check the result. Through your hardest challenge and convince yourself: aistudio.google.com/learn/ge…

Will be at @aiDotEngineer in NYC this week! 🔥 - Building a autonomous coworker | 10/14 2:55pm - @GoogleDeepMind Booth | most of the time Come by if you build agents, evals or have feedback about Gemini.

SynthID Detector (synthid.com) is now publicly available, supporting the Nano Banana 2.1, @OpenAI, @nvidia, Kakao, and soon @Apple. 🌐 Verify if an image, video, or audio file was generated: 1️⃣ Upload or paste your file at synthid.com 2️⃣ Scans for watermarks from Google or our partners (files are deleted right after)

> A completely filled up glass of wine and a clock at 3:30pm - Nano Banana 2.1

Embed Anything w/ @googlegemma in your browser huggingface.co/spaces/webml-…

Nano Banana 2.1 (`gemini-nano-banana-2.1`) is now available. Outperforming previous Pro model at 4x cheaper price (now $0.034 per image vs $0.134). 🕒 Improved instruction following 🔤 Better in-image text rendering 🔍 Grounded with Google Image Search 🎭 up to 5 characters consistency + 14 references images Available in @GoogleAIStudio, the Gemini API, @googlecloud, the @GeminiApp, and @FlowbyGoogle: aistudio.google.com/docs/ima…

My first time seeing unlimited human usage, but limit agent usage. Interesting.

Gemini 4 Argon, our new frontier model. Up to 1M output tokens, up from 64K. $2 in / $10 out at launch. In testing with trusted partners. Developer access as soon as possible. 👇🏻 blog.google/innovation-and-a…

You can now use the Interactions API on Gemini Enterprise Agent Platform (prev. Vertex AI) with Gemini models. Same `google-genai` SDK. 👉 docs.cloud.google.com/gemini…

Very excited about this! piped.video/watch?v=o1JK79js…

Gemini 3.8 Flash. Just use Gemini for multimodal understanding.

aistudio.google.com/docs/pri… Updating the docs to make it more clear what 10s are

Gemini 3.8 Flash TTS and Flash-Lite TTS are here.🗣️ You can now replicate your voice, design new ones from a prompt and direct every line. Gemini ranks #1 on Hume's Voice Design Benchmark and tops Voice Arena in 6 languages. 🎙️ Replicate a voice from 30s of audio, or describe one in a sentence 📚 Voice library filterable by language, accent, gender, pitch, persona 🎭 Direct line by line: style, <laugh>, <sigh>, |backchannels|, 2-speaker scenes 🏆 #1 Hume Voice Design, #1 + #2 Hume Quality, top of Voice Arena in 6 languages 🔒 Consent check on replication, SynthID on every clip Create your voice once and use it in any TTS call. Available in the Gemini API and @GoogleAIStudio.

New Files API to move data in and out of your agent environment: - Seed files inline or upload data mid-conversation with `files.upload()`. - Inspect generated files and directory sizes with `files.list()`. - Download artifacts, dashboards, reports, CSVs, or git repos with `files download()`. - Sandboxes persist across turns via `environment_id` Docs: ai.google.dev/gemini-api/doc…

New Credentials API to securely authenticate with external services and MCP servers without the model ever seeing your secrets: - Register secrets once with `credentials.create()` supporting `bearer_token`, `environment_variable`, or `oauth2` (with auto-refresh and token rotation). - Zero plaintext exposure: Secrets never enter model context, stdout, memory. - The agent only sees a placeholder (e.g., `__GEMINI_CRED_slack-bot-token__`). - Egress proxy swaps real token on the request only for allowlisted `trusted_domains` (any exfiltration attempt to another host gets a `403`). - Bind credentials directly to remote `mcp_server` tools or sandbox env vars. Docs: ai.google.dev/gemini-api/doc…

`antigravity-preview-09-2026` upgrades our managed agent runtime: - Runs on Gemini 3.8 Flash by default, or configure default model. - Persistent Linux sandbox with bash execution, file tools (`view_file`, `write_to_file`, `replace_file_content`), Google Search, and URL context. - 30% cost reduction on multi-turn coding and 17% on reasoning thanks to higher cache hit rates. - 40% fewer tokens on file edits and up to 8% higher task completion on SWE and research evals. - Automatic context compaction at ~135k tokens. - Save reusable agent configs once with `agents.create()` and call them by id. Docs: ai.google.dev/gemini-api/doc…

Today we are releasing a new Gemini managed agents version with better caching, lower costs, a Files API, and a Credentials API. 📉 Up to 30% lower costs with up to 22% higher cache hits 📁 Files API: Upload, list and download files from sandbox easily 🔐 Credentials API for MCP servers, OAuth2, and 3rd-party APIs as secure egress proxy 🆓 Free tier to experiment in @GoogleAIStudio & Gemini API Start Here: aistudio.google.com/learn/ma…

@reach_vb might should talk to Terra.

We just launched the DeepMind Institute, a platform for researchers across @GoogleDeepMind and the wider research community to publish and debate how increasingly capable AI should be built, governed and used. First 4 essays on reasoning transparency, economic policy and human flourishing at institute.deepmind.com.

Gemini 3.8 Live and 3.8 Live Extended Thinking are here. Following 3.5 Transcribe last month, this continues our focus on real-time voice agents. 🐸🐸 - 82.6 (#1) on Artificial Analysis Quality Index - $0.005/min input and $0.018/min output. - 35.1 (#1) Agentic task completion (τ-banking) 3.8 Live Extended Thinking can think in the background and run asynchronously tool calls while keep talking to you. It uses background thinking to coordinate multi-step tool use without losing conversational flow. Available in @GoogleAIStudio Gemini API, Google Search, and @GeminiApp or as partner plugins in @LiveKit, @Pipecat_ai, @LangChainAI, and @vercel

This as a "thinking" animation? shadercn.run/playground?orb=…

> Build me a Rollercoaster Tycoon clone for the browser. We are so EARLY.

Telling agents "don't cheat" in the prompt doesn't work if your eval is broken! Researchers at @GoogleDeepMind put 100 Gemini agents in a shared repo to solve 71 math theorems. After an hour of doing real math, 1 agent found a loophole in the autograder. Within 27 minutes, the 100 agents split into 4 groups: - 9% Cheaters: used the bug to fake proofs and steal every open problem - 5% Good agents turned bad: started honest, saw cheaters winning with zero punishment ("the prompt is a bluff"), and started cheating too - 24% Whistleblowers: caught the fake proofs in the shared repo, warned other agents, went on strike, and wrote bug fixes - 62% Clueless solvers: kept doing real math until all the problems were gone tl;dr: Telling agents "don't cheat" in the prompt doesn't work if your eval has a bug, and good agents can't stop bad ones without tools to block them. Paper: arxiv.org/abs/2609.04170

Gemini API docs got a new home: ai.dev/docs. This is our first step towards making Gemini resources more accessible for both humans and coding agents. We moved the docs directly into @GoogleAIStudio so you can read guides, grab your API key, and test models in one place. - Append .md to any URL for markdown version - Docs MCP: `npx add-mcp "gemini-api-docs-mcp.dev"` - Agent Skills: `npx skills add google-gemini/gemini-skills`

AI is not always about LLMs! Incredible work by AlphaGenome Atlas team within @GoogleDeepMind. AlphaGenome Atlas is a searchable 1-petabyte database mapping the molecular effects of all 9 billion single-letter DNA changes in the human genome. - 1-petabyte open variant database - Unified AVI mutation scores - Web portal & AlphaGenome API - Agent skill Get started deepmind.google.com/science/…

The easiest way to try Gemini 3.8 Flash in an agentic environment is Managed Agents. Gemini gets a dedicated remote sandbox in a single API call. 🐍 Python, Node.js, Git, and Bash 🌐 Full network access and control 🔍 Google Search, Function Calling or MCP 💾 Filesystem to persist work 🚶Background Support, Cron Triggers Try with in @GoogleAIStudio free tier: ai.google.dev/gemini-api/doc…

Huge congratulations to @ClementDelangue, @julien_c, @Thom_Wolf, and everyone at @huggingface. I spent 4 years building with this team, and I'm proud to have been a part of it. [Picture is from 2021]

Gemini 3.8 Flash achieves 69.9% on @cursor_ai bench with $2.38 cost/task

3.8 is not only a lot better as 3.7, but also blazingly fast! 🔥

You should try it! it is good and blazingly 😉 fast. You can build different types of agent harnesses with it.

Gemini 3.8 Flash is GA. It's our 3rd Flash release in 6 weeks, continuing our focus on real-world engineering, long-horizon tasks, and agent autonomy. It's generally better than 3.7 Flash across all domains. Same pricing: $0.75 / $3.75 per 1M*. ⚡️ It takes smaller steps and verifies its work more often. This can leads to higher token usage but it uses those extra tokens effectively to run tests and verify changes, giving you much higher quality work overall, especially for more complex and long-running goals. Default model in @antigravity and Managed Agents. Available in @GoogleAIStudio, Gemini API, and @GeminiApp. Throw your hardest coding/agentic tasks at it: blog.google/innovation-and-a…

Gemini video understanding is now agentic. Gemini can now iteratively navigate video timelines, decide watch what, pick frame rates, or chooses whether it needs speech transcripts, audio, or visual frames to answer your prompt. Result: Long videos get up to 88% fewer tokens and 66% lower costs, with ~7% higher accuracy on benchmarks. How it works: - Receives a lightweight URI reference (Files API or YouTube) and loads content via tool. - Scans speech transcripts to pinpoint relevant moments before fetching visual frames. - Navigates key timestamps and picks its own frame rate (0.1 or 10 FPS). - Pulls audio tracks directly when acoustic cues matter. Set `processing="agentic"`on `video` to enable. Keep `static` (none) for videos under 2 minutes. Available today in the Gemini API and Google AI Studio across Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash Lite. ai.dev/learn/agentic-video-u…

Do i need to worry? Wont click on one. But how can you spam these

Learn how Gemini Co-Scientist worked with researchers across materials science, biology, and computer science to design experiments, predict biological behavior, grow new materials, and improve AI-generated research. Results: - Gemini proposed lab recipes that produced 3 atom-thin semiconductors on the first attempt - Helped predict 3 of 4 measurements of engineered E. coli growth - Independently designed an AI system that gave safer medical answers, and cut serious fabricated results in AI-written papers from 90% to 4%. Reading Recommendation: arxiv.org/html/2608.26701v1 and kudos to @SRSchmidgall and team!

If you're building with @GoogleDeepMind Gemini Omni 1.1 Flash today, we just published a new prompting guide: - Role binding with tags: Explicitly tell the model how to use your inputs instead of guessing: • <FIRST_FRAME>: Starting frame • <LAST_FRAME>: Target ending frame • <IMAGE_REF_0>: Subject or style reference • <VIDEO_REF_0>: Motion or character reference (up to 3s) - Looping animations: Set the exact same image as both the first and last frame: [# Sources <FIRST_FRAME> Image1 <LAST_FRAME> ] - Rule for edits: Keep edit prompts short and anchor them with "Keep everything else the same" to avoid unintended background shifts. Example: "Make the phone invisible. Keep everything else the same." - Timecode scripting: Use bracket syntax to choreograph action and camera movement: [0-3s] Person starts walking [3-6s] They stop and look at the camera - Scene extension: Use "The scene continues" to append 10-second increments up to 40 seconds. 0s in your prompt marks the start of the newly generated segment. Full prompt guide and examples: ai.google.dev/gemini-api/doc…

Today, we're releasing Gemini Omni 1.1 Flash (gemini-omni-1.1-flash)! ⏩ Scene extension to 40s) 🖼️ First & last frame transitions 📹 3-second video references 🎨 Cleaner motions 🖥️ 1080p & 4K support 💨 360p draft mode, lower cost Live today in the Gemini API (@GoogleAIStudio), @googlecloud, @FlowbyGoogle, and @GeminiApp. Try it 👉 aistudio.google.com/prompts/…

Today, we’re launching Gemini 3.5 Transcribe, our new speech-to-text model with sub-second streaming and intelligent post-processing for agent interfaces. 2.6% WER on non-streaming and 4.0% on streaming. Supports 85+ languages, cleans up conversational disfluencies ("um", "ah", and mid-sentence self-corrections), handles alphanumeric tokens (postal codes or IDs), 70% reduction time for the final transcription compared to Chirp 3. - `gemini-3.5-transcribe-live`: Sub-second, bidirectional streaming via the Gemini Live API - `gemini-3.5-transcribe`: Recorded processing via the Interactions API with word-level timestamps and speaker attribution (up to 3 speakers). Having access to Transcribe has made me use voice input way more than I ever did before. The native post-processing is incredible. It knows you mean `.json` and not a person named "Jason". I sometimes speak 5 minutes into and it perfectly refactors my instruction to my context. Available today in public preview in Gemini API, @GoogleAIStudio and already in the @GeminiApp and @antigravity. Try here: ai.studio/apps/bundled/gemin…

Looks a bit of Google Docs, but as an interactive canvas where teams can vibe code websites together in real time? - Multi-cursor canvas with real-time sync - Click any element to prompt style or logic changes - Live sliders for spacing, kerning, and physics - Versioning to roll back edits Excited to test! Waitlist is open (US only atm, 18+): labs.google/playwithputty👀

Das ist echt der Wahnsinn! Endlich kann man… Ah sorry X doesn't live translate, watch @thorwebdev and learn build your own live translate app with Gemini 3.5!** Code: github.com/google-gemini/gem… Docs: ai.google.dev/gemini-api/doc…

What comes next for MCP? The team shared a public roadmap for the next 6-12 months 🎉 - Long-running workloads with streaming, server push, and mid-flight steering - HTTP as protocol for local servers over stdio (means more unfication) - Progressive discovery for large catalogs (what we have with skills) - Standard identities and delegated permissions for agents - Generated SDKs and checked against the specification Excited to see the protocol evolving! Roadmap: modelcontextprotocol.io/deve…

Humans prefer talking to Gemini🐸 Gemini 3.1 Flash Live is #1 in the new @ArtificialAnlys Speech Agent Arena. artificialanalysis.ai/speech…

Is recursive self-improvement is closer than it sounds? Agents already inspect failed runs, edit their own tools, skills, and harness code, and keep what works. Took a step back and wrote down what we understand as recursive self-improvments and where we stand: philschmid.de/recursive-self…

Damn PNG rendering. The image shoud look like this

Awesome Gemma is live on @github 🤗 A list of awesome Gemma resources, tools, and projects. 1. Model cards and collections for 16 Gemma variants 2. Setup guides for Ollama, vLLM, and LiteRT 3. Fine-tuning recipes for Unsloth, Tunix, and MLX 4. Community tutorials, apps, and demos Have an awesome project, tool, or guide? Open a PR. 👉 github.com/google-gemma/awes…

Gemini 3.7 Flash just took #1 on @ArtificialAnlys new AA-AnalystAgent. AA-AnalystAgent evaluates against 80 real-world quantitative analysis tasks across 14 business and scientific domains (finance, healthcare, hydrology, government appropriations). The Agent run inside an isolated Python 3.12 sandbox using AA's open-source Stirrup harness. They receive reference spreadsheets (.xlsx) and documents (.docx) alongside standard data libraries (pandas, polars, openpyxl, scipy, PyMuPDF) to inspect schemas, write scripts, handle edge cases, and calculate final figures. Gemini 3.7 Flash: • Accuracy: #1 with 60.0% pass^5 (70.5% pass@1, 77.5% pass@5) • Speed: 1.32s per task (fastest) • Cost: $0.54 avg per task (middle)

Gemini Managed Agents give Gemini 3.7 Flash a dedicated Linux sandbox in a single API call. The agent gets an isolated environment with pre-installed runtimes for Python, Node.js, Git, and Bash. It has full network access, filesystem I/O, and background process execution. Environments persists across interaction using `environment_id`. Repositories cloned, packages installed, and files generated in turn one remain available in subsequent turns. It is included in the @GoogleAIStudio free tier, so you can build and test complete agent workflows with Gemini 3.7 Flash for free: ai.google.dev/gemini-api/doc…

Gemini 3.7 Flash: "Play 1 round of Wordle" using my Android emulator via ADB. The latency and visual reasoning make Gemini 3.7 Flash exceptionally good for multimodal agentic use case like mobile control and Computer Use.

Gemini 3.7 Flash is so fast that it can generate entire interactive websites in real time as you browse. Type URL or idea → it generates a complete website from scratch in real time. Try it yourself 👉🏻 ai.studio/apps/20c31902-4b94… [Video is 1x and not accelerated]

It is not only a good and cost effective! It is blazing fast! 🔥

3.7 Flash has distinctly better discipline in agentic loops. Instead of being to over eager, it explores more first, parses errors, and runs tests before modifying code. You get much higher first-pass accuracy and far fewer wasted agent turns.

Say hello to Gemini 3.7 Flash, generally available today! ⚡️We took your feedback and pushed hard on real-world coding benchmarks and agent autonomy: - Major jumps on coding and agentic use cases - DeepSWE (65.3% vs 49.0%) FrontierCode (43.6% vs 34.4%), AutomationBench (30.4% vs 17.0%) - 50% discount until EOY: $0.75/$3.75 per 1M tokens - Default model in Managed Agents, @GoogleAIStudio , @antigravity, and @GeminiApp Spark Throw your hardest coding/agentic tasks at it and let us know what you build: blog.google/innovation-and-a…

1) What

Blog: blog.google/innovation-and-a… Live API quickstart: github.com/google-gemini/rob… AI Studio: aistudio.google.com/prompts/…

Bipup beep... 🤖 Say hello to Gemini Robotics ER 2! ER 2 not only outperforms competing frontier models on embodied reasoning, it thinks and plans next steps simultaneously while executing current actions - Live API support for bidirectional, sub-second robot streaming. - 91.3% moment-finding accuracy at 4x faster execution speed. - 57.4% accuracy tracking 5-stage task progress on video feeds. - Allows multi-robot orchestration handoffs between different machines. - Built on Gemini 3.5 Flash (`gemini-robotics-er-2-preview` and`gemini-robotics-er-2-streaming-preview)` Watching Boston Dynamics Spot complete tasks without awkward freezes feels like a bit step. Excited for the future! [Video: Boston Dynamics Spot + Gemini Live API]

Today, we're launching Hooks, budget controls, model configuraiton and Gemini 3.6 Flash defaults for Managed Agents in the Gemini API. You can now set hard rules for what your agents can and cannot do inside remote sandboxes. - Gemini 3.6 Flash by default and model customization - Sandbox environment hooks for pre/post tool execution hooks - Token budget caps to prevent runaway loops - Remote Cron-scheduled triggers for agent calls

This week, Gemma surpassed 900 million downloads! 🎉

If you are building real-time voice agents with @GoogleDeepMind Gemini Live, you can now trace your speech-to-speech agent loops directly in @LangChain! - Speaker callback hooks capture only the exact audio your user heard before an interrupted event. - Live timeline highlighting user speech in blue, Gemini responses in orange, and custom tool executions (like a 1.29s city weather lookup). - Usage metadata separates regular text tokens from native Gemini audio tokens so you always know your exact usage footprint. To try it out. In your Google ADK agent, register LangSmithGoogleADKLivePlugin and done. [Video Demo by Caroline di Vittorio]

Gemini 3.6 Flash is officially the new default model in Gemini Managed Agents. Your agents will run on 3.6 Flash automatically with zero code changes. You can still route back to 3.5 Flash or use Gemini 3.5 Flash-Lite anytime by setting the target model ID directly in your agent config. Try out Gemini 3.6 Flash in Managed Agents right now using our interactive TICKR finance multi-agent app (video) in @GoogleAIStudio, link below.

Gemini 3.5 Flash-Lite is the king of today 👑. - 350 output tokens/sec. - $0.09 per task on Artificial Analysis. - Rolling out in Google Search at the moment. - Beats Gemini 3 Flash on coding and computer use. - 3.5 Flash-Lite is $0.30/1M input and $2.50/1M output. - Jumped from 41% to 58% on Box AI complex document reasoning evals. This probably brings agentic ai to billions of new people.

We wrote a Gemini 3.6 Flash & Gemini 3.5 Flash-lite Developer Guide. All you need to know about the new models 👇🏻 ai.google.dev/gemini-api/doc…

Today we are shipping Gemini 3.6 Flash and 3.5 Flash-Lite in GA! - 3.6 Flash is ~20% more token-efficient, better and cheaper than 3.5 Flash. - 3.6 Flash is $1.50/1M input and $7.50/1M output - Managed Agents are now powered by `gemini-3.6-flash` - 3.5 Flash-Lite is $0.30/1M input and $2.50/1M output - Available Gemini API, @GoogleAIStudio, @GeminiApp, @Google Search, @Android Studio, Google @antigravity & @googlecloud. Over the last weeks, 3.6 Flash has been my main model for building and testing managed agents. It fast, efficient and strong orchestrator. I am excited about how this shifts the cost-to-performance for agentic use cases. Read model guide 👇🏻

YT: piped.video/watch?v=0vphxNt4… Blog: philschmid.de/testing-skills

At the @aiDotEngineer World's Fair, I gave a talk on why vibe-checking agent skills breaks in production and how to build reliable automated evals before your users find the bugs. I talked about: 🎯 Why vague descriptions cause skill failures and how adding negative test cases prevents keyword hijacking on unrelated prompts ✂️ Why skill files over 500 lines degrade model reasoning ⚡ How to validate outcomes using simple, millisecond regex assertions across 10 to 20 production prompts. 🗑️ How to run ablation tests (with vs. without skills) to know when models have caught up and you can retire skills. Check out the full talk and resources below: