langfuse / langfuse↗

Don't use n8n agents for business operations. Let me explain... I've been building AI agents for over 8 months now. I started with Flowise (great no-code platform), but slowly transitioned into n8n. I absolutely love n8n for it's powerful capabilities. In my opinion it's the best no-code AI agent builder and that's why I made it my primary AI agent building platform. Sure, it has a bit of a steep learning curve. But I think its well worth the investment learning it. So if it's so good, why don't I recommend using it for building production-ready agents for businesses. It might be tempting to sell these solution to other businesses or use it in your own business, but I would hold off on that. For a few reasons: 1. No-code platforms comes with inherent limitations. Although n8n is powerful, you can't "do what you want". If you are building an AI agent echo system for a client, there will comes a point where they'll make requests that are outside the scope of your platform. This could be a problem and will hinder growth in a market that is evolving at breakneck speed. 2. It's hard to handle errors gracefully. Errors will happen. It's not a question of if, it's a question of when. Although n8n has error handling built in, it's nowhere near as good as the options available with coded solutions. 3. It's hard to measure agent performance. One BIG thing people in the agent building space miss is creating systems for monitoring agents. This is important to troubleshoot and improving agent performance. If you are flying blind, you don't know what the agent is doing and why sub-optimal results are happening. By connecting your agents to something like Langfuse (langfuse. com), you can monitor every aspect of the agent and make informed decisions on how to improve its performance. Right now this is mostly available for coded solutions. 4. You don't truly "own" your agents. By building your AI agent infrastructure on someone else's platform you have created a co-dependency. If for whatever reason that platform is not available, your whole system is non-functional and there's nothing you can do about it. Where as if you coded your agents, you own the code and you are not dependent on a third-party. So does that mean you shouldn't use n8n? Not at all. I use n8n all the time. But I primarily use it for personal tasks. Since I'm not relying on it for important business operations, I feel comfortable with the potential downside of using n8n. It allows me to build quick agents for myself without having to worry about the complexity of code. I also use n8n to prototype agents before I commit them to code. This allows me to focus solely on logic and prompt optimization. So yes, there's a place for n8n. But just keep in mind, if you are building AI agents with scaling in mind, you should really consider coding agents in Python. If you are interested in building AI agents, vibe coding and building AI systems for you business check out Digital Alchemy. It's my newsletter where I share my thoughts and practical tips on building with AI. You can subscribe for free over at digitalalchemy. mba

Returned by a high-confidence repo query and contains a visible project phrase, but the exact URL, slug, or package name was not visible.

Day 5: Tracing & Observability—How Top LLM Teams Spot Issues Before Users Do Learning production LLM engineering in 30 days. Today: why end-to-end tracing is the real difference between a stable product and surprise outages at scale. Most developers rely on basic logs or uptime pings. In production, that’s not enough—what counts is tracing every step from the user’s request through vector search, LLM calls, agent chains, and back. What happens with best-in-class tracing? Every LLM request becomes a trace: you see when, where, and why things slowed or failed. You can directly tag each step with model name, input, latency, cost in USD, and quality metrics—catching anomalies in real-time. Teams use distributed tracing to debug hallucinations, replay problem chains, and prove exact costs per user/feature. Tech that’s winning in 2025: OpenTelemetry is the backbone standard—every major tool understands it. Pipe traces to Datadog( @datadoghq ), SigNoz , Dynatrace( @Dynatrace ) or Grafana ( @grafana ) Tempo for visual dashboards. Langfuse ( @langfuse ) and LangSmith are leading LLM-focused tracing platforms, integrating deeply with LangChain/RAG/agent stacks, offering detailed step-by-step workflow replay and cost tracking. New players like nexos.ai and established observability leaders like Dynatrace are expanding with dedicated LLM tracing modules. Why do companies invest early? Latency and cost spikes show up as trends, not one-off errors—tracing lets you catch runaway prompts or slow vector searches before users report issues. Many new platforms auto-detect anomalies, run LLM-based auto-evaluations, and alert you before problems go viral on social media. Takeaway: Your observability stack should trace every critical step—embedding, retrieval, LLM calls, and toolchains—with clear links to cost and quality. Companies getting this right move faster, debug easier, and cut headcount spent on fires. Tomorrow: How to build regression tests for prompt changes—so you never ship silent failures. Stick around. Every lesson gets you closer to real production wins that scale.

Returned by a high-confidence repo query and contains a visible project phrase, but the exact URL, slug, or package name was not visible.

ついに出ました、ユニファのLangfuse奮闘記！まだまだなかなかみないマルチモーダルでの奮闘記です。 #Langfuse LangfuseでマルチモーダルLLMアプリのトレーシングを行う際に工夫したこと - ユニファ開発者ブログ tech.unifa-e.com/entry/2025/… #生成AI #マルチモーダル #トレーシング

Returned by a high-confidence repo query and contains a visible project phrase, but the exact URL, slug, or package name was not visible.

Day 6: Prompt Regression Testing Every time you tweak your LLM prompt—no matter how small the change—you risk unexpected bugs, weird outputs, or user complaints. LLMs are extremely sensitive to wording, and without testing, tiny edits can break entire features. Here’s what actually works in production: Keep a curated set of real-life test prompts, each with its “gold standard” expected output. Before deploying a prompt change, run all tests and check if any key cases break or degrade. Tools like Langfuse( @langfuse ) and PromptLayer( @promptlayer ) automate version tracking and let you compare prompt performance, side by side and over time. Save all old prompt versions with a simple changelog. If something goes wrong, rollback instantly—no guessing. Want to be extra safe? Set up automated or scheduled “prompt checks” to catch silent failures fast, not days later. Big teams treat prompts like software code: they test, version, and only then release. If you’re shipping prompts without a safety net, it’s a matter of time before you sabotage yourself. How are you handling LLM prompt changes? Got a workflow to recommend? Reply below 👇 Tomorrow: Dashboards that monitor your LLM in real-time, so you catch problems before users do. Stick around for more production lessons.

Returned by a high-confidence repo query and contains a visible project phrase, but the exact URL, slug, or package name was not visible.

github.com/langfuse/langfuse

Contains the canonical GitHub repository URL.