Nhat Cuong Nguyen

The Kroxylicious project is very pleased to announce the release of Kroxylicious 0.16.0 [1]. See the Changelog [2] for a list of changes and summary of Deprecations, Changes and Removals. Apache Kafka 4.1.0 compatibility - The main change in this release is Apache Kafka 4.1.0 compatibility. The proxy will now relay Kafka 4.1 specific versions of the Kafka RPC and your filters can intercept them. Observability improvements - Thanks to the work of hrishabhg[3], the proxy now sports connection count gauges, allowing you to monitor downstream and upstream connection counts in realtime. [1] https://lnkd.in/gRw5a_SU [2] https://lnkd.in/gRUJ9h2J [3] https://lnkd.in/gxaPwviu

26

Claude Code and Claude Agent SDK are two great tastes that taste great together. The same "primitives" for working with Claude Code – Skills, Commands, Subagents – are you what employ when configuring your own agentic harness with the Agent SDK. So the lessons and best practices you learn while coding contributes DIRECTLY to the agentic system(s) you're architecting, and vice versa. You should probably be experimenting with multiple harnesses (for both coding and implementation). But if you can choose only one, it's hard to argue with Claude.

12

3 Comments

𝗜𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗶𝗻𝗴 𝟬-𝗦𝗵𝗲𝗹𝗹 — a 𝗨𝗻𝗶𝘅-𝗹𝗶𝗸𝗲 𝘀𝗵𝗲𝗹𝗹 𝗯𝘂𝗶𝗹𝘁 𝗳𝗿𝗼𝗺 𝘀𝗰𝗿𝗮𝘁𝗰𝗵 𝗶𝗻 𝗥𝘂𝘀𝘁. We created a lightweight, 𝗺𝗼𝗱𝗲𝗿𝗻 𝘀𝗵𝗲𝗹𝗹 𝗳𝗼𝗿 𝗟𝗶𝗻𝘂𝘅 𝗲𝗻𝘃𝗶𝗿𝗼𝗻𝗺𝗲𝗻𝘁𝘀—𝗻𝗼 𝗕𝗮𝘀𝗵, 𝗻𝗼 𝗲𝘅𝘁𝗲𝗿𝗻𝗮𝗹 𝘁𝗼𝗼𝗹𝘀—just pure system-level programming. 𝗪𝗵𝗮𝘁 𝗶𝘁 𝗱𝗼𝗲𝘀: Smart prompt with 𝗚𝗶𝘁 𝗯𝗿𝗮𝗻𝗰𝗵 + 𝗽𝗮𝘁𝗵 𝗖𝗼𝗹𝗼𝗿𝗳𝘂𝗹 𝗼𝘂𝘁𝗽𝘂𝘁 like Zsh (syntax highlighting, error hints) 𝗕𝘂𝗶𝗹𝘁-𝗶𝗻 𝗰𝗼𝗺𝗺𝗮𝗻𝗱𝘀 like cd, ls, cat, rm, echo, mkdir, etc. Handles signals (𝗖𝘁𝗿𝗹+𝗗), errors, and user input like a real shell 𝗪𝗵𝘆 𝗶𝘁'𝘀 𝗰𝗼𝗼𝗹: Written 𝟭𝟬𝟬% in Rust Re-implements 𝗰𝗼𝗿𝗲 𝘀𝗵𝗲𝗹𝗹 𝘁𝗼𝗼𝗹𝘀 Deep dive into terminal control, ANSI colors, input parsing, and more ... I worked on cat, rm, cd and some parsing tasks —focusing on file operations and graceful error handling while Amine Habchi made the project structure and worked on ls command in addition to cp , mv and parsing cmdline. Handling these features wouldn’t have been possible without abdelouahab khir who made a powerful parsing system as well as echo , mkdir , pwd tracking and other features. Check it out : https://lnkd.in/ev5fU4_b #Rust #Linux #Shell #SystemProgramming #OpenSource #CLI #TerminalTools

34

4 Comments

That's a wonderful way to promote it in the, probably, most conservative and pushback community. And I get the pushback - the wave of PRs lacking context made it increasingly hard to be a maintainer of any popular open source product. Well, let's see if the problem gets solved when AI is employed on the other side as well. Wonderful times we are living in!

1

1 Comment

New in /experimental mode in Copilot CLI: "Fleets" 🛸 Run `/fleet` to dispatch parallel subagents to implement your plan. The secret sauce here is a sqlite database per session that the agent uses to model dependency aware tasks and TODOs.

111

8 Comments

I wrote an article on how we translate user-generated content at scale at Mercari. It summarizes two years of work with various models, LLMs but not only, and how cost evolved, reducing it 100x over the period. I also go over non-AI features and user experience. https://lnkd.in/gbpQAQmX

39

2 Comments

For understanding the virtual thread state transitions... Specifically, thread pinning on Java 21 vs. Java 24 and when the thread waits on IO how the platform thread views the state compared against the virtual thread enabling interruption

3

What do we do at Allegro when we find promising technology? We battle-test it. 🛠️ More than six years ago, we built MBox, our in-house Server-Driven UI solution. Over time, as our ecosystem evolved, so did our expectations. We began exploring alternatives to our internal technology stack, looking for solutions that could offer the best of both worlds: the speed and flexibility of server-driven UI, combined with truly native performance and a modern developer experience. That search led us to Lynx, a recently released, open-source cross-platform framework designed around native rendering. On paper, it looked too good to ignore. Read the full story of why we’re exploring Lynx and how it integrates with our ecosystem on #AllegroTechBlog ➡️ https://shr.bi/wKWPKn5s

6

Been hands-on work with Amazon Q Developer recently, good base models and some interesting ideas, enjoying my time with it. 🎯 Always Plan Before You Act The best results come from plans, planing first, then execute. Tools like Cline nail this with explicit "plan" and "act" buttons - you can even push the planning phase to a cheaper model while reserving expensive compute for  execution, Amazon Q doesnt have this, but !always ask your AI to create a plan before taking action. Recent research papers back this up - agentic workflows with explicit planning phases significantly outperform back and forth conversation approaches. Also dont forget to ask it for questions! 👑 Context Is for Kings ( win if you know the quote ) Amazon Q really shines here with its profiles system, which I think beats tools like Gemini CLI hands down. Its secret sauce? Scope your context strategically: • Keep general contexts global across all profiles • Create focused, profile-specific contexts for specialized work • Always include business language in project contexts - your AI should speak your domain, not just code Pro tip: Got an internal library? Have your LLM generate context and bake it directly into projects that use it. This creates self-documenting, AI-friendly codebases. ⚡ MCP: Great Power, Great Responsibility Model Context Protocol integration is a game-changer, but use it wisely. Atlassian's MCP implementation is my go-to - it transforms how you work with tickets, documentation, and project context, it also enables some awesome workflows. But be selective about which external systems you connect. More isn't always better and there's is a vast quality cliff and thats not touching on the deep hole of MCP security. 🛠️ Command Line Precision Instruct your LLM to use command-line tools for anything requiring precision. Don't let your AI wing it on: • Code formatting and style enforcement • Package version management When your AI can directly invoke linters, package managers, and build tools, you get consistent, reliable results instead of "close enough" approximations. 💡And Finally ........ The Reality Check AI-assisted development isn't plug-and-play magic. There's a real learning curve, and success comes from combining thoughtful tooling with solid engineering practices.  The tools can bootstrap your workflow and handle routine tasks brilliantly, but they can't replace intentional implementation and good foundations like documentation, because guess what LLM needs to really shine documentation. But the engineer using the tool correctly remains the critical factor and theres a interesting conversation about a engineers who truly leverage these tools vs those that dont. I've been learning this tools, because as i've said before they are changing our industry, knowing what these tools will do to your engineering team is critical for the coming years or maybe .. tomorrow. Any Amazon Q tricks your want to share? #amazon #amazonq #llm #engineeringtooling

18

3 Comments

3 days ago I shared AgentLint — 10 rules to stop AI agents from drifting during long sessions. Since then I've shipped 4 major releases. Here's what changed: 10 rules → 42 rules across 7 packs AgentLint now auto-detects your stack and activates the right guardrails: 🐍 Python — bare excepts, unsafe shell, SQL injection, dangerous migrations 🎨 Frontend — accessibility checks (alt text, form labels, focus states, touch targets) ⚛️ React — loading states, empty states, lazy loading 🔍 SEO — metadata, Open Graph, semantic HTML, structured data 🔐 Security (opt-in) — blocks Bash file writes that bypass guardrails, catches data exfiltration attempts Full lifecycle coverage AgentLint went from 3 hook events to all 17 that Claude Code supports. New always-active quality rules: - Validates commit messages - Catches dead imports - Warns when error handling gets removed - Injects a self-review prompt at session end - Tracks session length and warns before things drift AGENTS.md support If your project uses AGENTS.md, AgentLint reads it and generates matching config automatically. Installation that just works I dogfooded a fresh install and hit the same PATH issues many developers face with pip on macOS. Fixed it — the resolver now probes 5 locations to find the binary. No more manual PATH hacking. The numbers 42 rules. 741 tests. 96% coverage. All running locally in milliseconds. pip install agentlint agentlint setup If you're running AI agents on production codebases, the question isn't whether they'll drift — it's whether you'll catch it before or after it ships. 👉 https://lnkd.in/eRMqpDHd #AI #ClaudeCode #DevTools #OpenSource #CodingAgents #DeveloperExperience #Python #AgentSafety

7

Need to rename an index in Elasticsearch? There’s no rename API, but here are practical workarounds. Check them out: https://gag.gl/R6bm3i #ElasticSearchLabs

4

We moved from RabbitMQ to AWS SQS. Reliability, not throughput, forced the decision. At our traffic levels, failures were silent. Backlogs growing quietly during peak traffic. On-call pain became our strongest signal. We asked a hard question internally. Can this system degrade gracefully at scale? RabbitMQ required constant operational tuning. Disk pressure. Broker health. Backpressure management. Too many sharp edges for our growth curve. SQS offered fewer operational failure modes. Managed durability with a simpler surface area. But reliability did not come for free. We redesigned consumers around idempotency. Visibility timeouts were treated as strict contracts. Lag became a first-class reliability metric. Retries were explicit, observable, and bounded. The result was a calmer system overall. Fewer incidents during traffic spikes. Monitoring became actionable. On-call was mostly uneventful. Engineering focus shifted back to delivery. At our scale, operational simplicity became reliability. What early reliability signal did you ignore and regret? #EngineeringLeadership #ScalableSystems #ReliabilityEngineering

42

6 Comments

AI-assisted solutioning makes me 𝕟𝕠𝕤𝕥𝕒𝕝𝕘𝕚𝕔. As part of Syren’s solutions team, I work with multiple models* to build scalable, feature-rich, service-oriented, and agentic supply-chain solutions. [* I often find myself returning to Claude Opus 4.5; read about Opus 4.6 this morning, will check it out] I am genuinely enjoying building closer to the speed of thought. But also pausing to observe and appreciate, between each session, the importance of context and intent. As much as I am training the models, I am also getting better at the art of planning, utilizing the tokens, multi-modal, and context refinement. And of course, validation of the solutions with functional experts and business leaders is imperative. That said, it’s hard not to feel a bit nostalgic about the "classic" days of software engineering. ▪️ PRs that encouraged deep review and healthy debate around standards and consistency (sometimes also abused by overly zealous engineers) ▪️ UX discussions that reflected just how seriously teams took user experience (sometimes at every level of leadership) ▪️ Thoughtful (and sometimes lengthy) conversations about dependency upgrades, technical debt, and long-term maintainability ▪️ Estimation and planning exercises designed to bring predictability and alignment to delivery. Agreed, sometimes they were futile ▪️ Strong architectural opinions formed through years of hard-earned experience Many forward-looking organizations are thoughtfully navigating this transition. They are embracing AI to accelerate delivery while still preserving the rigor, discipline, and institutional knowledge that made engineering teams effective in the first place. However, as things continue to evolve rapidly, the challenge (and opportunity) ahead is to continuously find the right balance. What parts of traditional engineering culture are worth preserving as AI reshapes how we build, and which should we finally let go? 🤔 #Syren #AIEngineering #SoftwareEngineering #EngineeringLeadership #GenAI #Claude #ClaudeOpus #AIAgent #Coding

36

3 Comments

Most verification folks won't agree with me, but the thing I find most exciting about efforts like this is the bug finding. Standard libraries are extremely highly leveraged places to find and fix bugs: you can fix millions of programs at the same time! Great to see ongoing progress on the Rust standard library verification effort.

165

6 Comments

Great article by Valery Frolov on how we created visibility and team-level usage attribution to drive a cost-aware mindset across the org. This was a key part of the process that led to a major cost reduction across our data platform

7

This time we will be musing about the importance of caching at the edge, how to think about TTLs, purging etc. Basically why Cache is King! 👑

3

Every engineering leader says they care about velocity. But velocity doesn’t die in the IDE. It dies in: Review queues Approval loops Dependencies across 3 other teams The fastest teams don’t code faster. They unblock faster.

1

Your CI/CD bill is not exploding because of “expensive runners,” it is bleeding out through flaky tests, over-parallelized jobs, and slow feedback loops that quietly tax every deploy This piece shows you how to redesign pipelines so you cut compute waste, shrink queues, and speed up developer flow without adding yet another approval gate or YAML religion Full breakdown in the comments. #DevOps #CICD #PlatformEngineering

8

1 Comment