I create content about deep learning | The DevRel and Deep Learning Dude

I don’t wanna hop on call to learn about or see a demo of what you’re building. Too busy hacking on VLMs, curating multimodal datasets, and contributing to the open source community. 

Harpreet Sahota 🥑

It was great to be back in Boston last week and host another awesome meetup!

Thank you for everyone who showed up, our awesome lineup of speakers for taking time out of their schedules for delivering such informative talks: Anku Rani Hanhui Wang Sage Simhon Sharut G. 

And for Voxel51 for sponsoring!

If you missed the event, checkout the comments for where you can find the recordings of the talks

🤖 Hacker-in-Residence @ Voxel51| 👨🏽‍💻 AI/ML Engineer | 👷🏽‍♀️ Technical Developer Advocate | Learn. Do. Write. Teach. Repeat.

No alternative text description for this image

i built a ComfyUI-inspired visual, node-based workflow editor for FiftyOne

now you can build dataset curation, analysis, and model evaluation pipelines by connecting nodes on a canvas, without writing code

give it a try and let me know what you think, completely open source and all you have to is `pip install fiftyone` and follow the instructions on the readme

everyone is hype about gtc, but the 3d vision conference is this week and i'm much hype for that

i made a repo and a dataset to explore the papers

checkout the repo here:
 https://lnkd.in/gRFz6Kb7

here's a dataset that you can use to explore the papers:
 https://lnkd.in/g2RA87fD

it's been a long, long while since i've done a live stream on linkedin, but it's happening in t-mius 8 minutes!

if you're interested in learning about how to work with video datasets using open source tools, open data, and open models...this is the session for you!

Join us on Feb 26 for Exploring Video Datasets with FiftyOne and Vision-Language Models. Register for the Zoom: https://lnkd.in/gCk7gGPi

Video is the hardest modality to work with — more data, more complexity, and annotation workflows that don’t scale. This hands-on session by Harpreet Sahota 🥑 shows how to understand what’s inside massive video datasets without manually watching thousands of clips.

You’ll learn how to:
🎬 Navigate and explore video datasets in the FiftyOne App and understand dataset structure
🔎 Compute embeddings with Qwen3-VL to enable semantic search, zero-shot classification, and clustering
📝 Generate video descriptions and localize events using vision-language models like Qwen3-VL and Molmo2
🧭 Visualize patterns and trends through embedding projections in FiftyOne
📏 Evaluate model outputs against Action100M’s hierarchical annotations

Perfect for ML engineers, data scientists, and computer vision teams working with large-scale video data who want practical workflows for exploration, curation, and debugging.

#computervision #machinelearning #ai #datascience #videoanalytics #visionlanguagemodels #mlops #opensource

Derek Van de Streek, Swati Bannore Jain and 11 other attendees

Feb 26 - Exploring Video Datasets with FiftyOne and Vision-Language Models

excited to be back in stuttart! if you're in the area and want to speak at the meetup on april 21st, send me a dm

Chief Community Officer @ Voxel51 | Startup Advisor

Are you based in the Stuttgart area? Consider speaking at the upcoming April 21 - Stuttgart AI, ML and Computer Vision Meetup: https://voxel51.com/speak

#computervision #ai #artificialintelligence #machinevision
#machinelearning #datascience #opensource #stuttgart

i've learned a lot from Eric Ma over the years, from the first time I saw one of his talks on principled data science workflows to talks he's given for my communities over the years, and his latest work on getting the most from coding agents.

check out his next talk, and support the SciPy community as well!

Together with my teammates, we solve biological problems with network science, deep learning and Bayesian methods.

As a data scientist, how much of your week is spent on work a coding agent could do?

I once planned multiple days of work for a machine learning experiment last year. This was an experiment that I had put off for an entire quarter because, as a team lead, I didn't have the time to block out the 3-4 afternoons that I thought I'd need. 

Instead, I described the problem to Claude Opus 4.5 via OpenCode, and it wrote the code, ran the training, found issues, switched strategies, and finished the whole thing in 20 minutes, complete with a written report drafted *with* figures referenced.

The data scientists who figure this out now will have an edge.

I'm teaching a workshop on how to do this flavour of agentic data science to raise funds for SciPy 2026's Financial Aid Program. You'll learn:

- Prescriptive prompting for experiments
- Project structure for agents
- Logging and verification strategies to enable automated control
- Report writing with AI
- Staying in control during EDA
- The minimalist version first

These are the same techniques I use in my personal projects and my teammates to compress data science timelines from days to minutes. I think you'll have a chance to learn something good while doing something good!

Details:

- Date: March 27 & April 3, 2026
- Time: 2:00-4:00 PM EST
- Venue: Google Meet (calendar invite will be provided)
- 40 spots available
- $1,000 minimum donation, paid directly to NumFOCUS (tax-deductible), student pricing: $200!

Video recordings available if you can't attend live.

Sign up here: https://lnkd.in/eFjc36Tb. I'm looking forward to seeing what you'll build with these techniques! 

P.S. If you can't make it but share it with others who have the means and would benefit from the workshop, I would deeply appreciate it!

Open article: Agentic Data Science Workshop | SciPy 2026 Fundraiser by agent-assisted-data-science.vercel.app

Agentic Data Science Workshop | SciPy 2026 Fundraiser

We just released a new AWS workshop that I’m really excited to finally share 👇
https://lnkd.in/eiNprTd7

This came from something I keep seeing across autonomy, robotics, and computer vision teams lately.

Most teams are not blocked by models anymore.
They are blocked by data and iteration speed.

So we wanted to show what a practical workflow actually looks like:
🚗 generate synthetic driving data
 🧠 auto label using foundation models
 🔧 fine tune a small vision language model
 ☁️ deploy and iterate using AWS SageMaker
 👀 inspect and improve everything inside FiftyOne

The goal was simple.

Build something real that teams could follow end to end.

Huge credit to Ryan Patrick Sweeney, Mark Vinciguerra, and Derick Thompson for driving this forward, along with great collaboration between Amazon Web Services (AWS), NVIDIA Omniverse, and Voxel51 🙌

Also excited that Voxel51 is now live on AWS Marketplace, which makes getting started a lot easier for teams working on computer vision and robotics workflows!

Personally, I think this is how a lot of Physical AI systems will get built going forward.

Faster loops between data, models, and deployment instead of long waiting cycles.

If you’re working on robotics, AV, or multimodal systems, I’d genuinely love to hear what you’re building.

More coming soon 🚀

#PhysicalAI #ComputerVision #Robotics #MultimodalAI #SyntheticData #AWS #MachineLearning #AutonomousSystems

Open article: GitHub - aws-samples/sample-finetuning-small-vlms-synthetic-data by github.com

Contribute to aws-samples/sample-finetuning-small-vlms-synthetic-data development by creating an account on GitHub.

GitHub - aws-samples/sample-finetuning-small-vlms-synthetic-data

I’m so happy to have my boi Wiz present at one of our meets!

AI @ NVIDIA - 🏗️ Build | 🚢 Ship | 🚀 Share

View Chris "The Wiz 🪄" Alexiuk’s profile

skills.md - you've seen it, you've probably used it (I assume if you're following this LinkedIn account, you're someone who uses tools like OpenCode/etc.

I'm excited to be yappin' about skills at an upcoming Voxel51 / Hugging Face virtual event with a line up that boggles the mind: 

Ben Burtenshaw, Daniel Ávila Arias, and AdonAI Vera.

My presentation will focus on something near and dear to my heart: Skills as documentation. 

The days of documentation aren't over - humans, and agents, still need to understand how things work - but they're changing quickly. Not everyone wants to read a thorough getting started guide that may be built for a specific system (and if it's not, it'll be ~2K words long) Sometimes, you just want to clone a repository - and get cooking. 

Skills help bridge that gap - context for LLMs, but tailored for a human experience. 

Can't wait for the event, come on out and ask any skill questions you have!

looking forward to hosting this workshop tomorrow, join me!

Join Harpreet Sahota 🥑 on Feb 26 for a virtual workshop to learn how to use Facebook's Action100M dataset and FiftyOne to build an end-to-end workflow - https://lnkd.in/gFvwbQni

What you'll learn:

* Navigate and explore video data in the FiftyOne App, filter samples, and understand dataset structure

* Compute embeddings with Qwen3-VL to enable semantic search, zero-shot classification, and clustering

* Generate descriptions and localize events using vision-language models like Qwen3-VL and Molmo2

* Visualize patterns in your data through embedding projections and the FiftyOne App

* Evaluate model outputs against Action100M's hierarchical annotations to validate what the models actually capture

By the end of the session, you'll have a reusable toolkit for understanding any video dataset at scale, whether you're curating training data, debugging model performance, or exploring a new domain.

***********

Want to build better computer vision models? FiftyOne is an open source toolkit from Voxel51 (our Meetup sponsor) that helps you curate datasets, evaluate model performance, visualize embeddings, catch annotation errors, and eliminate duplicate images—all in one place.

“pip install fiftyone” is all it takes to get started - https://docs.voxel51.com/

#computervision #ai #artificialintelligence #machinevision #machinelearning #datascience #opensource

AI product leader at Google, Author, and community builder making complex AI concepts simple for technical and business audiences. Published author of two books on GPT-3 and Neural Search. Lead and manage an active community of 350K+ AI developers across my social media channels and Unwind AI platform. Creator of Awesome LLM Apps, a #1 top-trending GitHub repository with 92k+ stars featuring practical AI Agents and RAG tutorials.

Building a community of 1M+ AI Developers at Unwind AI, an open AI education platform where 200k+ AI developers today learn to build with AI in just 3 minutes a day:
→ Daily AI news that actually matters
→ Code-first tutorials on AI Agents, RAG and LLMs
→ Open-source AI Agent and RAG apps you can use today with full code implementation

Author of books on: 
→ "GPT-3: The Ultimate Guide To Building NLP Products With OpenAI API"
→ "Neural Search - From Prototype to Production with Jina"

Previously:
→ Built Tenstorrent's developer community from scratch & launched first generation of AI hardware products from AI Accelerator cards to AI Desktop Workstations.
→ Built AI systems for Wall Street hedge funds @Arcesium (DE Shaw)
→ Led data initiatives at 256 Network
→ Tech Leaders Fellow @Plaksha

I share daily tips and tutorials on:
→ Building production-ready LLM apps
→ Implementing RAG systems
→ AI Agents and opensource agentic frameworks

🤝 Let's connect if you're:
→ Building with AI and need practical guidance 
→ Looking for ready-to-use LLM implementations
→ Want to stay ahead of AI developments and grow your career in AI

Follow me for daily AI insights and tutorials on AI Agents, RAG and LLMs.

Shubham Saboo

This open-source project just made vector databases optional for RAG 🤯

No embeddings. No chunking. No similarity search.

Meet PageIndex, a vectorless, reasoning-based RAG system that retrieves like a human expert.

It just hit 98.7% accuracy on FinanceBench, outperforming traditional vector RAG by a significant margin.

Vector search finds what's semantically similar, but similarity is not relevance. 

When you're working with financial reports, legal filings, or technical manuals, you need reasoning, multi-step logic, and context across sections. A similarity score can't give you that.

PageIndex thinks differently.

It reads your document and builds a hierarchical tree index, like a table of contents but optimized for how LLMs actually think. 

Then it reasons through that tree to find what's actually relevant, the same way a human expert flips through a 200-page annual report.

What that looks like in practice: 
• Documents stay in their natural structure, no artificial chunking
• The LLM navigates the tree using reasoning, not pattern matching
• Every answer comes with page and section references
• No opaque embeddings, no "vibe retrieval"

Works on SEC filings, legal documents, technical manuals, and academic papers. Anything that needs precision and exceeds your context window.

Vector database RAG was the best we had. This is what we actually needed.

Link to the GitHub repo in the comments.

Senior AI Product Manager @ Google | Awesome LLM Apps (#1 AI Agents GitHub repo with 103k+ stars) | 3x AI Author | Community of 350k+ AI developers | Views are my Own

Someone just dropped an open-source AI voice cloning studio that runs locally on your computer 🤯

Clone any voice from a few seconds of audio. Generate speech. Keep everything private.

Meet Voicebox. The local, free alternative to ElevenLabs.

Here's what makes this different:

No cloud. No subscription. No voice data leaving your machine.

Upload a voice sample. Get a voice profile. Generate speech with that voice. All running locally.

But this isn't just a voice cloner. It's a full studio.

Multi-track timeline editor. Trim and split clips. Mix multiple voices into conversations. Create podcasts, narratives, dialogues. DAW-like features for voice synthesis.

It runs on the open-source Qwen3-TTS model from Alibaba. 
Near-perfect voice cloning from just a few seconds of audio. 
Natural prosody, emotion, and cadence.

On Apple MacBook's, MLX backend gives you 4-5x faster inference with native Metal acceleration.

It also exposes a full REST API for:
• Game dialogue systems
• Podcast production pipelines
• Accessibility tools
• Voice assistants
• Content creation automation

Available now for macOS and Windows. Linux coming soon.

No Python install required. No cloud dependency. No limits.

Your voice data stays yours.

Link to the GitHub repo in the comments.

Senior AI Product Manager @ Google | Awesome LLM Apps (#1 AI Agents GitHub repo with 101k+ stars) | 3x AI Author | Community of 350k+ AI developers | Views are my Own

Someone just open-sourced an entire operating system for 24/7 AI agents 🤯

And it fits in a single 32MB binary.

Meet OpenFang. Built from scratch in Rust. 137K lines of code. 

Every agent framework today waits for you to type something. You prompt it, it responds, you prompt again. 

OpenFang runs autonomous agents that work for you on schedules, 24/7, without you sitting at a terminal.

They call these "Hands." Pre-built autonomous capability packages that run independently.

Here's what that looks like in practice:
• Researcher Hand wakes up, cross-references sources, evaluates credibility, generates cited reports
• Lead Hand runs daily, discovers prospects matching your ICP, scores them 0-100, delivers qualified leads before you check email
• Collector Hand monitors targets continuously with change detection, sentiment tracking, and knowledge graph construction
• X and LinkedIn Hand manages your account with an approval queue so nothing posts without your OK
• Browser Hand automates web workflows with a mandatory purchase gate so it never spends your money without confirmation

7 Hands bundled. You can build your own with a HAND.toml file.

Cold start under 200ms. 40MB idle memory. 27 LLM providers. 53 built-in tools plus MCP and A2A support. 
16 independent security systems including WASM sandboxing, Merkle hash-chain audit trails, and prompt injection scanning.

Already running OpenClaw? One command migrates everything.
openfang migrate --from openclaw

The gap in the agent ecosystem wasn't another orchestrator. It was an operating system that treats agents as first-class processes, not chat sessions.

Link to the GitHub repo in the comments.

This is what a one-person AI agent company looks like in 2026 🤯

6 AI agents. 20 cron jobs. 0 human employees.

Every role is a folder. Every job description is a markdown file. No standups. No Slack. No payroll.

Just a directory on a Mac Mini that runs the whole thing.

I am sharing the actual file tree.

Each agent has a name, a personality, and one job.

Monica is the Chief of Staff. She coordinates everything.
Dwight runs research sweeps three times a day.
Kelly drafts tweets from the research.
Rachel writes LinkedIn posts.
Ross handles engineering and code reviews.
Pam turns the daily intel into a newsletter.

Each agent is a folder. Each SOUL file is a job description, personality, and operating manual in one file.

shared-context is the company brain. THESIS md is what we believe. SIGNALS is reference intel. FEEDBACK-LOG is every style correction I have ever given.

memory is continuity. Daily logs. Long-term memory. Agents wake up fresh every session and load context from these files. They get better over time not because the model improves but because the memory gets richer.

The coordination layer is the filesystem itself. Dwight writes DAILY-INTEL. Content agents read it. No API calls between agents. No message queues. No orchestration framework. Just files.

Scheduling is cron. Dwight runs first because everyone depends on his output. A heartbeat system checks if anything missed its window and forces a re-run. Self-healing. No babysitting.

I wake up, open Telegram, and the night shift is done. Research summary waiting. X drafts ready. LinkedIn post drafted. Ten minutes over coffee.

Everyone has access to Claude, GPT, Gemini. The alpha is this directory. 

The SOUL files. The memory. The feedback logs. The weeks of context stored in plain text.

That system compounds every day. And nobody else has yours.

This open-source agent orchestrator lets you run 30 AI agents in parallel 🔥

Spawn agents. Monitor from one dashboard. Merge their PRs.

Here's the problem it solves:
One AI coding agent in a terminal? Easy
30 agents across different issues and PRs? Coordination nightmare

Without orchestration, you're manually creating branches, starting agents, checking if they're stuck, and reading CI failures. 

On top of that, forwarding review comments, tracking which PRs are ready, and cleaning up when done. That doesn't scale.

Here's how it works:
ao spawn my-project 123

That's it.

Isolated git worktree. Feature branch. tmux session. Claude Code with full issue context. All from one command.

Agent works autonomously. Reads code. Writes tests. Creates PR.

CI fails? Agent reads the logs and fixes it.
Reviewer requests changes? Agent addresses them.
PR approved and green? You get notified to merge.

You only get pulled in when human judgment is needed.

The plugin architecture is clean:
• Agent-agnostic: Claude Code, Codex, Aider
• Runtime-agnostic: tmux, Docker, Kubernetes
• Tracker-agnostic: GitHub Issues, Linear

Eight swappable slots. Every abstraction is replaceable.

This is the pattern. You don't run agents. You orchestrate fleets.

One dashboard. Parallel workstreams. Automated feedback loops.

Link to the GitHub repo in the comments.

"Delete your CLAUDE.md files" 🤯

That's the hottest take in the AI agents community right now.

1000+ posts on X today debating it.

And most of them are missing half the study.

ETH Zurich just dropped a new research testing whether CLAUDE or AGENTS MD files actually help coding agents.

138 real GitHub issues. 4 coding agents. one question: do context files help agents ship better code?

the success metric was simple: did the agent produce a working patch?

not tokens. not speed. working code.

two findings came out of it.

most people are only sharing one.

auto-generated context files (the ones from running /init) hurt task resolution by 0.5-2% and bump inference cost by 20%.

that part is true.

but human-written ones improved performance by 4%.

same study. different number.

the reason auto-generated files hurt is kind of funny once you see it.

agents are obedient. they follow whatever's in the file.

so when a generated file adds unnecessary requirements, the agent does more testing, more file traversal, burns up to 22% more reasoning tokens, and still produces worse results than if you'd given it nothing.

now here's where it gets interesting.

when researchers stripped all existing documentation from the repos, auto-generated files went from hurting to helping by 2.7%.

repos with good documentation don't need a generated context file.

the agent was already going to find that information on its own.

/init just repackages what's already there.

what actually works is telling the agent things it cannot figure out from your codebase itself.

which test runner to use.
which package manager you prefer.
non-obvious repo-specific commands.

the researchers tracked actual tool usage to prove this.

mention `uv` in your context file and the agent uses it consistently.

don't mention it and the agent almost never reaches for it on its own.

so no, don't delete it.

if you ran /init and committed whatever came out, that's the thing to delete.

write a few lines yourself instead.

the study doesn't say CLAUDE.md is useless.

it says auto-generated, bloated ones are.

just stop letting the loudest take win over the actual data.

AI agents can now write their own Agent Skills in Google ADK 🔥

No manual SKILL authoring. No redeployment. The agent generates new capabilities on demand.

Here's the problem every agent builder hits eventually.

You build a library of skills. SEO checklist. Blog writer. Code reviewer. Your agent loads them on demand and performs well.

Then a user asks for something you didn't anticipate. Security vulnerability review. API documentation. Compliance audit.

The agent has no skill for it. You stop. Write a new SKILL manually. Redeploy. Repeat forever.

Meta skills in Google ADK eliminate that entire loop.

You give the agent one skill called skill-creator. Its instructions explain how to write valid SKILL files. Its references contain the agentskills spec itself and a working example.

When the agent hits a request it has no skill for, it does this automatically:

• Loads the skill-creator instructions
• Reads the spec via load_skill_resource
• Reads the example skill via load_skill_resource
• Generates a complete, spec-compliant SKILL

Valid YAML frontmatter. Structured instructions. Proper references/ directory. Ready to load in the next session.

The generated skill isn't locked to ADK either. It works in any agent that follows the agentskills spec. Claude Code, Gemini CLI, GitHub Copilot, Cursor. All of them.

A skill library you maintain is useful. A skill library that grows itself is a different category entirely.

Link to the GitHub sample and docs in the comments.

Google is back with another season of Advent of Agents 🔥

150,000 of you joined us for Season 1 in December. That was just the beginning.

We heard you. You want more. So we built more.

Season 2 is now live. 31 days. One agent topic per day. 

From zero to production-ready AI agents. 100% free.

Join now: https://lnkd.in/dsCF8_Fb

Here's what drops every day in March:
⚡ One new agent concept or pattern
⏱️ One kata you can run in under 5 minutes
💻 Copy-paste code that actually works
📖 Direct links to the ADK docs

No theory. No fluff. Just build and ship.

What you'll learn this season:

Week 1: Build your first ADK agent with Gemini 3.1 Pro. Add MCP servers, long-term memory, and secure code execution.

Week 2: Wire agents together. 7 multi-agent patterns, each one copy-pasteable. Sequential pipelines, parallel fan-out, human-in-the-loop, and more.

Week 3: Go deep. Real-time voice with Live API, multimodal agents, context management, and cross-framework orchestration with A2A.

Week 4: Harden everything. Evals, Model Armor security, production observability, and deploying to Vertex AI Agent Engine.

The archive is permanent. Miss a day? Go back. Need that pattern from Day 8? It's there.

Season 1 archive is still live if you missed the foundation.

Day 1 starts today. Let's build AI Agents this March.

Google just dropped Gemini 3.1 Flash-Lite 🔥

And it is the model to build always-on multimodal AI agents.

I just built an open-source memory agent with Google ADK that runs 24/7 as a background process.

Feed it anything. Text, images, audio, video, PDFs. It reads, connects and consolidates everything while you sleep.

Here is how it works:

1️⃣ Ingest
Drop any file into the inbox folder. The agent auto-extracts structured info using Gemini's multimodal capabilities. Entities, topics, summaries, importance scores. Supports 27 file types.

2️⃣ Consolidate
Every 30 minutes, it replays and connects your memories. Like the human brain during sleep. It finds cross-references, generates insights and compresses related information. No vector database. No embeddings. Just an LLM that thinks.

3️⃣ Query
Ask it anything. It reads all memories + consolidation insights and gives you a synthesized answer with source citations.

Why Gemini 3.1 Flash-Lite? It’s low latency, negligible cost for 24/7 operation, and smart enough to find connections and synthesize answers.

Most AI agents have amnesia. They process when asked, then forget everything.

This one remembers, connects and evolves.

This is Day 3 of Google Advent of Agents Season 2.

Missed Day 1 and 2? The archive is live. Catch up anytime.

Link in comments.

Google Colab just became a tool any AI agent can control 🔥

Not a new UI. Not a notebook feature. Programmatic access to the full Colab environment.

Your agent can now create notebooks, write code, execute it, and organize results. All in the cloud. All without touching your local machine.

Every developer using AI agents hits the same wall. The agent generates code in your terminal, then you copy-paste it into Colab to actually run it. Debug something? Copy it back. Visualize data? Switch tabs again.

That context switch kills flow. The Colab MCP Server removes that entire loop.

Here's how it works:
Connect any MCP-compatible agent to Google Colab. Gemini CLI, Claude Code, or your own custom agent.

Tell it: "Load the sales dataset and forecast next month's numbers."

Then watch.

Your agent creates the notebook. Adds markdown cells explaining its methodology. Writes Python to load pandas and matplotlib. Executes the code. Generates visualizations. Formats the final report.

All happening live in your Colab notebook. You can jump in at any point to inspect state or take over.

What your agent can now do inside Colab:
• Create and structure .ipynb files with markdown and code cells
• Execute Python in real time with full Colab compute
• Install dependencies that aren't in the base image
• Rearrange cells to build a logical, readable flow

No more copy-pasting between terminal and browser. 
No more running heavy compute on your local machine. 
No more agents bottlenecked by your hardware. 

Works with any MCP-compatible agent.

Link to the GitHub repo in the comments.

We just released an enterprise guide to multi-agent systems 🔥

It covers everything: how to build, scale, and govern AI agents in production.

Most enterprises focus 90% of their effort on building the agent.

And 10% on everything that actually determines if the agent system survives in production.

Here's what those problems actually are (and how enterprises are solving them):

📌  Your agents need an identity 
• Not a login. 
• A real, managed identity with granular permissions and a full audit trail. 
• Without this, you have no accountability. No compliance. No way to debug what went wrong at 2am.

📌  MCP is becoming the universal adapter
• Instead of building custom integrations for every data source, MCP lets your agents connect to BigQuery, AlloyDB, Postgres through one standard layer. 
• 100+ partners already support it. This is moving fast.

📌  A2A protocol solves the hardest problem in multi-agent systems
• Getting agents built by different teams, on different frameworks, to actually talk to each other. 
• A2A does this via a simple "agent card" (a JSON file) that advertises what the agent can do. Any A2A-compliant agent can discover it and send tasks to it. 
• Box built this. Geotab built this. The ecosystem already has 100+ companies.

📌  Agent payments are now a real protocol
• AP2 handles authorization, authenticity, and accountability when an agent makes a purchase on your behalf. 
• This is early but important. Agents that can transact autonomously need a completely different trust layer than agents that just answer questions.

📌  The eval loop is what separates agent demos from products
• ADK has built-in observability and an Example Store for few-shot steering. 
• You can improve agent accuracy on specific tasks without retraining the model.
• This is huge for teams running 10+ agents at once.

The 3-phase framework by Google Cloud that you should know: Build. Scale. Govern.

Most teams only invest in phase 1.

The ones shipping agents in production are spending equal time on all three.

Link to the full guide in the comments.

Agent Skills are now live in Google Agent Development Kit 🔥

And they completely change how your agents load knowledge.

Right now, every agent builder does the same thing. Cram everything into the system prompt. Compliance rules. Style guides. API docs. Troubleshooting steps. One giant wall of text.

User asks about SEO? Your agent is still dragging around research templates, code standards, and troubleshooting procedures it will never touch.

10 tools in a system prompt means ~10,000 tokens burned before the agent even starts thinking.

Agent Skills replace that entire pattern.

Your agent gets a menu instead of a manual. Short descriptions of everything it knows how to do. ~100 tokens per skill.

It reads the query. Picks the right skill. Loads only those instructions.

Needs a reference doc? Fetches it on demand. Not upfront. Not "just in case."

Three levels of progressive disclosure:
• L1: skill names and descriptions. Always loaded. The menu.
• L2: full instructions. Loaded only when the agent picks that skill.
• L3: reference files, templates, assets. Loaded only when instructions call for them.

10 skills with progressive disclosure? ~1,000 tokens per call.

Same capabilities. 90% less context wasted.

ADK's SkillToolset generates the tools automatically. Wire up your skills and the agent handles the rest.

Built on the open standard Agent Skills spec adopted by Claude Code, Gemini CLI, GitHub Copilot, Cursor, and OpenAI Codex.

Available now in ADK Python v1.25.0.

System prompts were the best we had. Skills are what we actually needed.

Link to the GitHub sample and docs in the comments.

JUST SHIPPED: Open Source OpenClaw Agent Plugin for Google Vertex AI Memory Bank 🔥

Tell one agent your preferences. Every agent remembers. Across sessions. Forever.

Here's the problem with OpenClaw's built-in memory.

It's per-agent and per-session.

You tell Agent Monica your deployment runs on us-central1. Agent Dwight has no idea. You restart a session. Gone.

Your agents don't share context. And they don't carry memories across conversations.

This plugin fixes both.

It connects your OpenClaw agents to Vertex AI Memory Bank. Fully managed. No vector database to run. No embeddings to maintain.

Here's how it works:
Before every turn, the plugin retrieves relevant memories via semantic search and injects them into context.

After every turn, it extracts facts from the conversation and stores them.

Not raw conversation logs. Extracted facts.

"User prefers us-central1 for deployments" not a 400-token transcript about debugging a region config.

When a new fact contradicts an old one, it updates automatically. Your repo went from 95K to 97K stars? Memory Bank resolves it. No duplicates. No stale data.

What makes this different:
• Auto-recall: Before each turn, relevant memories are retrieved via similarity search and injected into context
• Auto-capture: After each turn, the last message pair is sent to Memory Bank for fact extraction and storage
• Noise filtering: Short exchanges are automatically skipped. Few-shot examples teach Memory Bank what to extract vs ignore
• File sync: Workspace files are automatically synced to Memory Bank with hash-based change tracking
• Agent tools: Search, forget, correct, and inspect memory stats directly from conversation

No vector DB. No local database. No infrastructure to monitor.

Your agents should compound knowledge over time, not reset with every session.

Link to the GitHub repo in the comments. 

Send this repo link to your OpenClaw agent and watch it install itself.

Introducing the Google ADK Tools and Integrations Ecosystem 🔥

AI agents are only as powerful as the systems they can interact with. 

An agent that can 'think' is great, but an agent that can manage code repositories, trigger workflows, or query databases is actually useful.

Today, we're expanding Agent Development Kit with 30+ integrations from the best developer platforms, so your agents can interact with the real world with just a few lines of code.

Your ADK agents can now:

💻 Manage code, test APIs, and execute scripts in secure sandboxes
- Daytona, GitHub, GitLab, Postman, Restate

📋 Track projects, manage issues, and search documentation
- Asana, Atlassian, Linear, Notion

🔍 Query databases and perform semantic vector search
- Chroma, MongoDB, Pinecone

🧠 Persistent semantic memory across conversations
- GoodMem, Qdrant

📊 Monitor, debug, and optimize agents in production
- AgentOps, Arize AX, Freeplay, MLflow, Monocle, Phoenix, W&B Weave

🔗 Connect to 200+ external apps and services
- n8n, StackOne

🤖 Access AI models, datasets, and research papers
- Hugging Face

💳 Handle payments, invoices, and subscriptions
- PayPal, Stripe

🔊 Generate speech, clone voices, and create audio content
- Cartesia, ElevenLabs

📧 Automate email workflows and manage agent inboxes
- AgentMail (YC S25), Mailgun

The best part? No refactoring your agent. 

Just configure the tool or plugin, and ADK handles the rest.

Link to the full blog in the comments.

This open-source AI agent turns docs into training videos with consistent characters 🤯

Not 8-second clips. Full long-form videos. Same avatar. Same scene. Every shot.

Video generation models peak at 8-second clips. That's great for demos. Useless for actual training content.

The moment you try to chain clips together, your character changes hairstyle, your background shifts, and your "consistent brand video" looks like it was filmed across three continents.

This agent system fixes that with a multi-agent coordinator pattern built on Google ADK.

Here's how it works:

• An Orchestrator agent receives your request and dispatches work to two specialists.

• Script Sequencer takes dense documentation and converts it into natural speech. Then segments the script into 8-second chunks.

• Video Agent generates each clip using the Orchestrator's instructions and a reference frame. That reference frame is what keeps your character and scene consistent across every single chunk.

What's under the hood:
• Gemini 3.1 Flash Image (Nano Banana 2) for image generation
• Veo 3.1 for video generation via MCP Server
• Vertex AI Agent Engine Sessions for state management
• Deploy in a few commands on Google Cloud Run

Upload your documentation. Attach your reference images. Hit enter.

The agent converts, segments, generates, and stitches. You get a training video.

No manual editing. No scene-by-scene prompting. No consistency headaches.

This is the Coordinator pattern applied to something beyond text. Agents that don't just write. Agents that produce.

Link to the GitHub repo in the comments.

Senior AI Product Manager @ Google | Awesome LLM Apps (#1 AI Agents GitHub repo with 102k+ stars) | 3x AI Author | Community of 350k+ AI developers | Views are my Own

90% of AI agents forget everything the moment you close the tab.

No memory of your preferences. No recall of past conversations. Nothing.

Here's the pattern that fixes this. It's called persistent semantic memory, and it takes about 10 lines of code.

How it works:
1. A plugin silently captures every message and converts it into a vector embedding
2. Before each new turn, a similarity search pulls the top 5 most relevant memories from ALL past sessions
3. Those memories get injected directly into the system prompt

Result: your agent starts every session already knowing the user. Dietary restrictions, coding preferences, project history, all of it.

Two ways to add this to Google ADK agents today:

GoodMem Plugin 
• Attaches at the App layer via callbacks
• Saves user messages, agent responses, and file attachments
• Retrieves relevant context every turn with zero manual effort

Qdrant MCP Server
• Store and find tools let your agent decide what to remember
• Works for RAG, code search, knowledge bases
• Runs locally with Docker or on Qdrant Cloud

The difference between a toy agent and a real one isn't the model.

It's whether it remembers who it's talking to.

That's Day 5 of Google Advent of Agents Season 2.

New hands-on tutorial every day through March.

Follow along at adventofagents.com

Senior AI Product Manager @ Google | Awesome LLM Apps (#1 AI Agents GitHub repo with 99k+ stars) | 3x AI Author | Community of 350k+ AI developers | Views are my Own

taplio

The Battle

Harpreet Sahota 🥑

Shubham Saboo

Followers

Posts

Engagements

Likes

Performance ScoreHarpreet Sahota 🥑's LinkedIn Performance

Performance ScoreShubham Saboo's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Harpreet Sahota 🥑

Shubham Saboo

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Harpreet Sahota 🥑's LinkedIn Performance

Performance ScoreShubham Saboo's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreHarpreet Sahota 🥑's LinkedIn Performance