Carlos Santana Roldán

Agents are the new apps. And they just went multiplayer. The software engineering landscape is undergoing a seismic shift. Google just dropped Firebase Studio. And it might just replace half the AI builder startups out there. You’ve probably seen tools like Lovable, bolt.new, Vercel's v0, or Cursor. They let you build AI-powered apps without writing much code. Describe what you want — boom — they create UIs, workflows, and automation for you. It’s fast. It’s smart. It’s changing how software is built. Now imagine that… backed by Google. What is Firebase Studio? Firebase Studio is Google’s new end-to-end app builder for AI-first products. It combines: ✅ UI generation ✅ Agent workflows (like AgentSpace) ✅ Firebase backend (auth, storage, hosting, etc.) ✅ Google’s best Gemini models AgentSpace is like a command center for AI agents. It lets you create multistep workflows, connect APIs, and manage how agents behave. Great for product teams prototyping internal tools, automations, or AI assistants. But it’s still pretty early-stage. Right now, Firebase Studio and A2A (Agent-to-Agent) are two separate moves from Google — but they are strategically aligned, and integration is very likely coming. You need to manage agents, instructions, tools, and sometimes code. Why this matters for product managers? → The barrier to building software is decreasing. → With Firebase Studio, a product manager can do what used to take a team. The AI agent is becoming the new app. Not just chat. Agents that do stuff: send emails, fetch data, run logic. Google wants to standardise agents across platforms. Their new Agent-to-Agent protocol (A2A) means your agents could soon talk to each other — across apps, tools, maybe even companies. Enter MCP: The USB-C for AI Anthropic’s Model Context Protocol (MCP) is an open standard that enables developers to build secure, two-way connections between their data sources and AI-powered tools. MCP is a standardised protocol designed to streamline the way AI models communicate with external systems, eliminating the need for custom integrations. With so many advancements happening rapidly, it’s challenging to keep up and navigate them within the realities of development. Yet, it has never been more exciting to be in the software engineering space. #AI #FirebaseStudio #ProductManagement #Innovation #TechTrends #MCP #A2A

16

1 Comment

"Context engineering” is just prompt engineering with added context. The LLM looks at two things: 1. Prompt = Guard rails, Format, Instructions, etc... 2. Context = What documents/code/chat history to include, Summary, etc... How is it different from plain prompt engineering? The queries are heavily augmented here, with chat history summary and other relevant documents. Think of it as RAG + structured prompts. Calling it "engineering" is a stretch. It's too broad and domain-specific to be a science. If you are a software engineer looking to learn more about AI engineering, try this structured course at InterviewReady: https://lnkd.in/gAFcCkRj Cheers! #ContextEngineering #LLMs #AI

55

5 Comments

🚀 FLUX just took another leap — and the generative crowd is buzzing. Black Forest Labs has officially rolled out FLUX.2, and it’s not just “a new version.” It’s a full, multi-tier release designed for everyone—from enterprise pipelines to indie experimenters. 𝙃𝙀𝙍𝙀"𝙎 𝙏𝙃𝙀 𝙉𝙀𝙒 𝙇𝙄𝙉𝙀𝙐𝙋 👇 ✨ 𝗙𝗟𝗨𝗫.𝟮 [𝗽𝗿𝗼] – The flagship. Enterprise-grade quality, licensed for commercial work, built for teams that need reliability at scale. ▪️ 𝗟𝗶𝗰𝗲𝗻𝘀𝗲: Commercial / Proprietary license via API or enterprise agreement ▪️𝗨𝘀𝗮𝗴𝗲: Full commercial rights for production, advertising, client work, etc. ▪️𝗡𝗼𝘁𝗲𝘀: Requires a paid license or API usage. Not open-source. No redistribution. ⚡ 𝗙𝗟𝗨𝗫.𝟮 [𝘀𝗰𝗵𝗻𝗲𝗹𝗹] – The speedster. Light, fast, local-friendly, perfect for personal workflows and rapid iteration. ▪️ 𝗟𝗶𝗰𝗲𝗻𝘀𝗲: Non-commercial license (for personal and research use) ▪️𝗨𝘀𝗮𝗴𝗲: - Personal experimentation - Local/private projects - Non-commercial prototypes ▪️𝗡𝗼𝘁𝗲𝘀: - Cannot be used for commercial work or monetized outputs. - Redistribution prohibited. 🔧 𝗙𝗟𝗨𝗫.𝟮 [𝗱𝗲𝘃] – Open-weights for researchers and builders who want to dig into the internals. ▪️ 𝗟𝗶𝗰𝗲𝗻𝘀𝗲: Non-commercial open-weights license (similar to FLUX.1 dev) ▪️𝗨𝘀𝗮𝗴𝗲:  - Research - Academic work - Personal experimentation - Model exploration, fine-tuning (with non-commercial restrictions) ▪️𝗡𝗼𝘁𝗲𝘀:  - No commercial use permitted. - Cannot deploy in any revenue-generating product or service. 📦 𝗙𝗟𝗨𝗫.𝟮 [𝗸𝗹𝗲𝗶𝗻] (coming soon) – Fully open-source under Apache 2.0. ▪️ 𝗟𝗶𝗰𝗲𝗻𝘀𝗲: Apache 2.0 (fully open-source) ▪️𝗨𝘀𝗮𝗴𝗲: - Unlimited commercial use - Model redistribution - Fine-tuning and derivative works - Integration into SaaS, products, client pipelines, etc. ▪️𝗡𝗼𝘁𝗲𝘀: This will be the first FLUX 2.x variant that allows true commercial freedom without licensing fees. 🔍 𝗪𝗵𝗮𝘁’𝘀 𝗻𝗲𝘄 𝘂𝗻𝗱𝗲𝗿 𝘁𝗵𝗲 𝗵𝗼𝗼𝗱? FLUX.2 is built on a latent flow matching architecture that fuses generation + editing in one unified system. ⭐ 𝙡𝙚𝙖𝙧𝙣 𝙢𝙤𝙧𝙚 𝙖𝙗𝙤𝙪𝙩 𝙙𝙞𝙛𝙛𝙪𝙨𝙞𝙤𝙣 𝙫𝙨 𝙛𝙡𝙤𝙬 𝙨𝙮𝙨𝙩𝙚𝙢𝙨 𝙞𝙣 𝙘𝙤𝙢𝙢𝙚𝙣𝙩𝙨 ⭐ This unlocks some powerful capabilities: 🟢 Combine multiple reference images and guide the model with surgical precision 🟢 Faster inference with noticeably improved consistency 🟢 4MP native output that looks studio-ready without post-processing 🟢 Strong VLM backbone via Mistral’s AI-3 24B model + rectified flow transformer magic This is a big workflow upgrade 😇 Freepik already confirmed integration is rolling out, so expect it to land across partner platforms shortly. 🙌 Kudos to the Black Forest Labs team for continuing to push the boundary on high-fidelity image generation. This release is big. The future is Agentic 🤖 📁 Try it now - links in comments 👇 #flux #blackforestlabs #comfyui #flowmatching #monks

103

10 Comments

Building a GenAI app? Check out Shalini Goyal Don’t just plug in a model - design it to scale, adapt, and evolve. Here’s your blueprint for future-ready GenAI systems. 👇 1. Modular Architecture Separate UI, orchestration, models, and storage to swap parts independently. Use LangChain or LlamaIndex to build pipelines. 2. Context Engineering Layer system prompts, memory, and retrieved knowledge to optimize generation. Use chunking and summarization to stay efficient. 3. Retrieval-Augmented Generation (RAG) Connect vector DBs like Pinecone or Weaviate and use hybrid search (dense + keyword) for domain-specific relevance. 4. Low-Latency Design Cut load times and delay using model distillation, quantization, and async I/O. 5. Agent-Based Systems Use CrewAI, AutoGen, or LangGraph for task decomposition and tool execution via specialized sub-agents. 6. Tool & Plugin Integration Enable LLMs to run code, hit APIs, or use external tools through OpenAI function-calling or LangChain routing. 7. Streaming & Feedback Improve experience with real-time streaming via WebSockets and user feedback for continuous refinement. 8. Memory Management Support both session and long-term memory using Redis, Postgres, or vector DBs for persistence. 9. Smart Deployment Use K8s or serverless runtimes (like AWS Lambda) to deploy GenAI apps with dynamic scaling. 10. Observability Track usage, hallucinations, and prompts using tools like LangSmith or WhyLabs for LLM monitoring. [Explore More In The Post] Here’s the takeaway? Good GenAI apps aren’t just about prompts, they’re engineered for performance, adaptability, and scale.

21

1 Comment

Here is Gemini CLI’s October 27th weekly update for v0.12.0 - 🎉 New Partner Extensions: - 🤗 Hugging Face extension: Access the Hugging Face hub. gemini extensions install https://lnkd.in/g2_Ek8Ua - Monday.com extension: Analyze your sprints, update your task boards, etc. gemini extensions install https://lnkd.in/gZbqPThi - Data Commons extension: Query public datasets or ground responses on data from Data Commons. gemini extensions install https://lnkd.in/g8dkRGsg - Model Selection: Choose the Gemini model for your session with /model. (abhipatel12) - Model Routing: Gemini CLI will now intelligently pick the best model for the task. Simple queries will be sent to Flash while complex analytical or creative tasks will still use the power of Pro. This ensures your quota will last for a longer period of time. You can always opt-out of this via /model. (abhipatel12) - Discussion: https://lnkd.in/gRPKHduA - Codebase Investigator Subagent: We now have a new built-in subagent that will explore your workspace and resolve relevant information to improve overall performance. (abhipatel12 and silviojr) - Enable, disable, or limit turns in /settings, plus advanced configs in settings.json (silviojr) - Explore Extension: Users can now open the extensions page in their default browser directly from the CLI using the /extensions explore command. (JayadityaGit) - Configurable Compression: Users can modify the compression threshold in /settings. The default has been made more proactive (scidomino). - API Key Authentication: Users can now securely enter and store their Gemini API key via a new dialog, eliminating the need for environment variables and repeated entry. (galz10) - Sequential Approval: Users can now approve multiple tool calls sequentially during execution. (joshualitt) https://lnkd.in/gcGSDux8 🧵 #GeminiCLI #Gemini #AI #OSS #HuggingFace #Monday

181

11 Comments

Roo Code lets you mold your #AI coding agent into what you actually need. Brandon Mathis from This Dot Labs built custom modes for documentation, pair programming, writing specs, and improving his prompts. If you're serious about agent workflows (which you have to be), this is worth a read.

13

4 Comments

Why doesn’t SwiftUI support SVG? It’s 2025, and we still need to add a third-party library just to render an SVG in a SwiftUI app. That’s wild. Instead of native support, we’re stuck: - Converting to PDF (if we want vector scaling) - Embedding PNGs (and losing scalability) - Or wiring up SVGKit, Macaw, or similar just to handle a format the web has handled for decades. Meanwhile, SwiftUI handles: ✅ PNG, JPEG, HEIC, PDF, even TIFF ❌ But not SVG or animated GIFs (natively)

5

Claude revenue went up from less than $1B to $4B ARR in 7 months. Here is where the money is coming from. Most people think it's from Claude subscriptions. They're dead wrong. Only 10-15% comes from direct subscriptions. Every time you use Cursor, Windsurf, or Microsoft Copilot, you're burning Claude tokens. Every time an enterprise uses Amazon Web Services (AWS) Bedrock and Google Vertex AI to create solutions or every time a Manus AI comes up, they are all using Claude tokens. Code generation eats 10-50x more tokens than chat. That's where 75% of their $333M monthly revenue lives. While OpenAI chases consumer buzz, Anthropic built something different: Their developer-led growth engine is scaling faster than any SaaS model I've seen. There are three patterns I can see: 1. API-first beats sales-first - API first models do not need sales teams. Build something which doesn’t need a large sales team to scale. 2. High-token workflows = high-margin streams - Code generation, data processing, content creation - all token-heavy 3. Leverage existing distribution - AI Coding tools, AWS Bedrock, Google Vertex AI remove procurement friction. Their sales team are selling your product This isn't just the gold rush analogy everyone uses for NVIDIA. It's more like the railroad boom of the 1860s. Everyone needed to ship goods. Few built the rails. Anthropic built the rails. The question for every CTO or Entrepreneur - Can you think of any other widespread need which can be converted into a separate API based SaaS product?

28

1 Comment

Are you manually setting up mock data for Compose previews? There's a faster way! If you haven’t tried it yet, Gemini in Android Studio can help automate Compose preview generation. Simply right-click on a composable and navigate to Gemini > Generate "<composable>" Preview or Generate Compose Previews for this file. It will then show a different view where you can accept Gemini’s suggested Compose preview code or reject the proposed changes.   If you’ve been using this feature, I’d love to hear what you think!

25

1 Comment

If you’re a recruiting firm thinking about building your own AI stack: Here’s my rule: Build if you have a real CTO + continuous engineering capacity. Buy if you don’t. “Vibe coding” with a part-time dev is how you spend: $100k waste hundreds of hours and end up with a fragile system you now have to maintain forever Vibe coding ≠ reliability. And reliability is the product. Ask yourself: Who’s on call when it breaks? How are you managing PII and permissions? Are you prepared for legal/compliance realities (including data broker registration in some cases)? What happens if LinkedIn bans a profile for unapproved tooling? If you’re building, I’ll swap war stories anytime.

27

10 Comments

If you haven't used an agentic CLI tool yet, do yourself a favor by installing Node.js and running these commands in your Terminal: npm install -g @google/gemini-cli gemini Gemini CLI still trails Claude Code in user-friendliness and agentic handling, but it's large context window and web search capabilities make it a joy for use. Gemini CLI + Wispr Flow is one of my favorite ways to brainstorm, research, and build on ideas.

10

𝗟𝗟𝗠 𝗶𝗻𝘁𝗲𝗴𝗿𝗮𝘁𝗶𝗼𝗻 𝘁𝗼 𝗼𝘂𝗿 𝗮𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻 𝘁𝘂𝗿𝗻𝗲𝗱 𝗱𝗲𝗯𝘂𝗴𝗴𝗶𝗻𝗴 𝗶𝗻𝘁𝗼 𝗮 𝗻𝗶𝗴𝗵𝘁𝗺𝗮𝗿𝗲. 𝗛𝗲𝗿𝗲'𝘀 𝗵𝗼𝘄 𝘄𝗲 𝗳𝗶𝘅𝗲𝗱 𝗶𝘁 𝗶𝗻 𝟱 𝗱𝗮𝘆𝘀 𝘄𝗶𝘁𝗵 𝘇𝗲𝗿𝗼 𝗻𝗲𝘄 𝘁𝗼𝗼𝗹𝘀. 𝗧𝗵𝗲 𝗣𝗿𝗼𝗯𝗹𝗲𝗺: A single user request triggers 200+ LLM calls. Which one failed? Where did context break? Our MTTR became as high as 7 hours. As a Startup, we cannot afford to have our reputation at stake. 𝗧𝗵𝗲 𝗦𝘁𝗮𝗿𝘁𝘂𝗽 𝗗𝗶𝗹𝗲𝗺𝗺𝗮: Everyone suggested: LangSmith, LangFuse, Helicone. - Evaluating tools = 2-3 weeks - Every new tool = Learning curve + cost + vendor management - We don't even know what we need yet 𝗙𝗶𝗿𝘀𝘁 𝗣𝗿𝗶𝗻𝗰𝗶𝗽𝗹𝗲𝘀 𝗤𝘂𝗲𝘀𝘁𝗶𝗼𝗻𝘀: What exactly do we need to debug? Input sent, response received, post-processing results, and relationships between 200+ calls. The insight: Not a logging problem. It's a 𝘁𝗿𝗮𝗰𝗶𝗻𝗴 𝗽𝗿𝗼𝗯𝗹𝗲𝗺. Each LLM call is stateless. Our application maintains context. Same applies to observability. 𝗧𝗵𝗲 𝗦𝗼𝗹𝘂𝘁𝗶𝗼𝗻: We already had Grafana Loki, Tempo, and OpenTelemetry. The breakthrough: Connect all LLM calls under one parent span. 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲 (𝗿𝗲𝗳𝗲𝗿 𝗱𝗶𝗮𝗴𝗿𝗮𝗺): 𝟭. 𝗔𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻 𝗟𝗮𝘆𝗲𝗿 OpenTelemetry emits logs and traces 𝟮. 𝗟𝗼𝗴𝘀 (𝗚𝗿𝗮𝗳𝗮𝗻𝗮 𝗟𝗼𝗸𝗶) Captures each LLM interaction: - trace_id & span_id - Message sent & response received 𝟯. 𝗧𝗿𝗮𝗰𝗲𝘀 (𝗚𝗿𝗮𝗳𝗮𝗻𝗮 𝗧𝗲𝗺𝗽𝗼) Parent span connects all related calls: - "Analyze PDF" → Factor 1 (Profit) → Factor 10 - Each factor → LLM Call 1... Call 20 Shows serial vs parallel execution, timing 𝟰. 𝗢𝗯𝘀𝗲𝗿𝘃𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗨𝗜 (𝗚𝗿𝗮𝗳𝗮𝗻𝗮) Query by span_id → Complete journey with all logs 𝗧𝗵𝗲 𝗥𝗲𝘀𝘂𝗹𝘁𝘀: - MTTR: 7 hours → 30 minutes - Timeline: 5 days(2 design + 2 dev + 1 infra) - Cost: $0 in new tools - Complexity: Zero new vendors 𝗞𝗲𝘆 𝗗𝗲𝘀𝗶𝗴𝗻 𝗗𝗲𝗰𝗶𝘀𝗶𝗼𝗻𝘀: 𝗪𝗵𝘆 𝗧𝗿𝗮𝗰𝗲𝘀 𝗢𝘃𝗲𝗿 𝗟𝗼𝗴𝘀? Logs show events. Traces show relationships. For 200+ interconnected calls, relationships matter. 𝗪𝗵𝘆 𝗘𝘅𝗶𝘀𝘁𝗶𝗻𝗴 𝗦𝘁𝗮𝗰𝗸 𝗢𝘃𝗲𝗿 𝗡𝗲𝘄 𝗧𝗼𝗼𝗹𝘀? Existing stack solved 80%. New tools solve 100% but cost 5x in time, money, and complexity. 𝗪𝗵𝘆 𝗣𝗮𝗿𝗲𝗻𝘁-𝗖𝗵𝗶𝗹𝗱 𝗦𝗽𝗮𝗻𝘀? Mirrors how our application maintains state. One request → Multiple factors → Multiple LLM calls. 𝗦𝘁𝗮𝗿𝘁𝘂𝗽 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗟𝗲𝘀𝘀𝗼𝗻𝘀: Don't: Add new tools before understanding the problem Do: Define what you need first (saved 3 weeks) Don't: Evaluate 5 tools when you don't know what you need Do: Use existing stack, learn what's missing, choose deliberately Don't: Wait for the "perfect" solution Do: Ship 80% solution in 5 days, iterate based on production issues We'll eventually need specialized AI observability tools. But not today. Tomorrow, we'll know exactly which tool solves which problem. That's 𝗦𝘁𝗮𝗿𝘁𝘂𝗽 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴: Choose simple over perfect solution. Solve for 80%. Iterate to 100%.

28

2 Comments

Don't believe the BS that you can use Claude Code for free. Ollama recently made their API compatible with Claude Code. Many creators quickly jumped on the opportunity to farm engagement with the hook: "You can now use Claude Code for free!" My thought? Claude Code without Opus 4.5 is not Claude Code. Period. But this is exciting news. Not because I can use Claude Code for free, but because I see an opportunity to optimize costs by delegating easier tasks to local LLMs. The key question is: what tasks can local LLMs handle? I tested out 7 local LLMs. In this post, I will explain the BS and share my 1st experiment. -- Why the BS? Claude Code has been praised as one of the best AI tools by its users, not only for coding but for many other tasks. But the price feels steep to many. The $20/month plan is not enough for any serious work. You need to at least subscribe to Max 5x ($100/month). Many heavy users, including me, subscribe to Max 20x ($200/month). It’s a steal. But still, many were eager to try it but aren't ready to pay. Ollama's recent announcement means you can buy a Mac, a Strix Halo, or a GPU and use Claude Code for "free" with local LLMs. It’s appealing, as it is a one-time investment, and you can use the machine for other purposes. Creators are leveraging this opportunity to farm engagement. But the reality? Claude Code without Opus 4.5 is not the same Claude Code we praised. Local LLMs are far less intelligent. -- But for those who understand the difference, we see an opportunity to optimize costs by delegating some easier tasks to local LLMs. I'm interested in finding out what tasks local LLMs can handle. This is my typical flow when using Claude Code. This is for coding, but I have similar flows for marketing and content creation. 1. Research and planning 2. Create PRD and implementation plan 3. Break plan into bite-sized tasks 4. Implement + review with reflection pattern 5. Final review with agents 6. Final review and QA by me Based on my quick tests, we can forget about asking local LLMs to do research and planning. All of them failed at a simple instruction: "Visit https://learnparrot.ai/ and tell me about the website." So, I think the most viable use cases would only be (4) — implementation + review loops. While it looks like a very narrow use case, it is where we burn a lot of tokens. So I think it is worth a try. The main selection criteria for this will be instruction-following capability. One very common task is to refer to code samples or templates to code a new feature or page. This is a good test of instruction-following. So, I crafted my first test: - Used Opus to create HTML that I can screenshot as a LinkedIn carousel to display info for each model. - Turned one of the pages into a template and deleted the rest. - Asked each LLM to refer to the template to code its own page, given its specs. Swipe the carousel to see the results. Who would you hire? #LocalLLM #ClaudeCode #VibeCoding

20

10 Comments