Theo Browne

💡 Over the past two days, I’ve had a surprising number of founders, infra leads, and creator teams reach out with questions about production multi-agent systems. I’m not pivoting into consulting, I’m just getting a lot of inbound messages this week, and I wanted a simple way to respond. Here are the five failure modes I see most often when systems move from demos to real users: - Cold-start cascades → one stalled agent collapses the entire chain - BigQuery or vector DB latency → blows the 300ms budget instantly - Voice synthesis drift → load breaks timing and pacing first - Memory inconsistency → agents contradict each other under pressure - Concurrency spikes → systems that look fine at 5 users fail at 40–300 These issues don’t appear in demos, only when real users start firing agents at the same time. Since several teams asked for help this week, I put together a small page outlining the areas where I can support builders who are scaling: 👉 https://lnkd.in/gSqkhrmH If you’re approaching a December launch or hitting one of these bottlenecks, feel free to reach out. Happy to help where I can. #MultiAgentSystems #AIInfrastructure #SystemDesign #VertexAI #LatencyEngineering #MLOps #DistributedAI #RealTimeAI #VoiceEngineering #ScalableAI

2

This week my team dove into a cal.com open source codebase for a class project — and found something worth fixing. We opened the browser's accessibility inspector on the live booking page and checked what a screen reader actually announces to blind users navigating the calendar.                                                          Here's what it says:                                                          "4. Button. 4. Button. Disabled, button. 1:30pm, button."  No month. No year. No "unavailable." No "Book." Just numbers floating in silence.                                                                We also found that disabled calendar dates fail the legal color contrast minimum — by a small enough margin that most people would never notice visually, but enough to fail WCAG 2.1 AA compliance under the European Accessibility Act.               So we built a fix. Two pull requests, one week, three people:                               - One line of code fixes the contrast    - Three small label additions make the calendar fully navigable by screen reader                                                                                    Submitting to the Cal.com open source repo this week.                                                                #OpenSource #Accessibility #CalCom #BuildInPublic #Pursuit

26

1 Comment

Before: 3+ hours on a plan. Exhausting. After: 5 bullets. ~15 minutes. PRD ready to build! In June, I gave a talk at Agentic Builders Collective about using the grill-with-docs agent skill by Matt Pocock and the shaping agent skill by Ryan Singer in a highly customised way. The end outcome was a PRD + implementation plan that could be used to build production-grade software by feeding it into AI harnesses like Codex or Claude Code. Over the next few days, I received feedback from a few of you who had tried out the approach. The through line was that it took way too long: All took at least 3 hours, and one even took over 8 hours! They were looking for a faster alternative. I have been reflecting on that feedback a lot. Production-grade software does require investing in an intense amount of planning upfront, there's no escaping that. Especially if a team of humans are NOT going to be building that over the course of several weeks or months, where there are lots of opportunities to discover/rectify issues in the plan. Instead, we're actually going to give it to generative AI, let it rip, and in a matter of minutes the whole thing will be done. Conundrum! New problems require fresh thinking, and new solutions. But many times, the new solutions can be built on ideas from old ones. In this case: Prototyping! Instead of planning a large/complex project up front, simply build a prototype of that software. This used to be expensive, as even simple software used to take significant time/effort to build. Now, generative AI makes it so fast/cheap, that I think prototypes should be far more commonly used. Use your prototype to inform you of the true shape and specs of the thing you want to build. Here's how to use this new agent skill: - Install the prototype-plan skill - Create a REQS.md file in your project directory - Insert 5 bullet points about what you want to prototype - Prompt to invoke the skill: /prototype-plan - Let it interview you for up to 15 mins - It will output PRD.md and IMPL.md, which you can use in your implementation step - Build your prototype! (Agent skill linked in comments ↓) How many hours did your last planning cycle burn? Save this for your next greenfield project.

22

7 Comments

🔥 Curiosity about useState drove me into the world of Fiber and reconciliation I started learning about the useState hook and how the React team actually built this thing. It looks like everything is tightly connected with Fiber. The hook logic lives inside Fiber, and the whole idea of custom hooks is just to reuse logic anywhere you want. The only rule is the function name must start with use*, but inside we just write normal JavaScript logic. When I checked how useState is implemented, I noticed they use Fiber heavily. Then I fell into the rabbit hole of Fiber and reconciliation. Very interesting stuff. Before React 16, they used something called stack reconciliation. It was fully synchronous, no priorities, no way to pause, and everything executed through the JavaScript call stack (LIFO), which caused a lot of problems for big apps. Then Fiber came and completely changed this. Instead of relying on the JS call stack, React now has its own Fiber nodes. Whenever you update something, React creates a separate work-in-progress Fiber tree in the background (a detached copy of the current one). After all the work is done, React commits it, and the Fiber root starts pointing to the newly committed tree. That’s basically how modern React avoids blocking the UI and handles priorities. People still say React “Virtual DOM” because it’s easy and they don’t understand Fiber architecture or how React really works internally after React 16 btw latest version is (19.x). Pretty interesting right ? #ReactArchitecture #DeepDive #TechExplained #ReactFiber #ReactReconciliation

3

1 Comment

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

21

10 Comments