https://letscode.hu/en/blog Software architecture, AI engineering, and delivery — from letscode.hu. en https://letscode.hu/en/blog/p/you-dont-need-an-agent-for-that https://letscode.hu/en/blog/p/you-dont-need-an-agent-for-that Tue, 11 Aug 2026 00:00:00 GMT A team spent three weeks building an agent to triage support tickets, then replaced it with a text classifier - faster, cheaper, more predictable. Most agent-shaped ideas are tasks with a mapping already sitting in the data, waiting to be learned instead of reasoned about turn by turn. https://letscode.hu/en/blog/p/ai-didnt-make-you-5x-faster https://letscode.hu/en/blog/p/ai-didnt-make-you-5x-faster Thu, 30 Jul 2026 00:00:00 GMT An agent generated a feature in minutes, then sat idle waiting for CI runners. AI made writing code cheap, but that was never the same as making software cheap - and the gap is where the promised 5x disappears. https://letscode.hu/en/blog/p/architect-wont-be-here-when-bill-comes-due https://letscode.hu/en/blog/p/architect-wont-be-here-when-bill-comes-due Tue, 14 Jul 2026 00:00:00 GMT The person who mandates "everything serverless" is systematically the one least exposed to the consequences. The decision outlives the decider, and the misalignment between who decides and who pays is the real defect - not the tech. https://letscode.hu/en/blog/p/best-engineer-barely-writes-code https://letscode.hu/en/blog/p/best-engineer-barely-writes-code Sat, 20 Jun 2026 00:00:00 GMT A team lead almost put his best engineer on a performance plan for low output - until his manager asked one question. When generating code is nearly free, output stops measuring value, and judgment becomes the work no dashboard can see. https://letscode.hu/en/blog/p/ai-model-independence https://letscode.hu/en/blog/p/ai-model-independence Mon, 15 Jun 2026 00:00:00 GMT An export-control order disabled a frontier model for everyone - even the US customers it never targeted. The fix is not a model from the right country, but an architecture that can swap models without a rewrite. https://letscode.hu/en/blog/p/spec-driven-development-waterfall-trap https://letscode.hu/en/blog/p/spec-driven-development-waterfall-trap Sun, 07 Jun 2026 00:00:00 GMT Spec-driven development solves real problems with vibe coding, but writing a long, confident spec up front quietly brings back the waterfall assumption. The fix is to iterate the spec, not just the code. https://letscode.hu/en/blog/p/ai-generated-adrs https://letscode.hu/en/blog/p/ai-generated-adrs Mon, 01 Jun 2026 00:00:00 GMT AI can produce polished ADR-shaped text, but it cannot weigh your real decision drivers. Use it to expose assumptions and blind spots before the team owns the choice. https://letscode.hu/en/blog/p/testing-in-ai-age https://letscode.hu/en/blog/p/testing-in-ai-age Mon, 11 May 2026 00:00:00 GMT AI can generate tests fast, but speed often amplifies brittle, implementation-focused test suites. The real gain comes from behavior-focused tests that survive refactors. https://letscode.hu/en/blog/p/you-are-not-building-airplanes https://letscode.hu/en/blog/p/you-are-not-building-airplanes Tue, 05 May 2026 00:00:00 GMT Reliability does not improve by writing fallbacks for every uncertain state. Critical points need redundancy; everywhere else needs simplicity. https://letscode.hu/en/blog/p/overengineering-vs-underengineering-compliance-nightmare https://letscode.hu/en/blog/p/overengineering-vs-underengineering-compliance-nightmare Sun, 26 Apr 2026 00:00:00 GMT A tenant deletion endpoint looked like five lines, but multi-region data, retention rules, and auditability turned it into a business-critical workflow design problem. https://letscode.hu/en/blog/p/worth-solving-before-architecture https://letscode.hu/en/blog/p/worth-solving-before-architecture Mon, 20 Apr 2026 00:00:00 GMT Teams often start with stack and architecture decisions too early. First validate whether the problem is worth solving now, then design the smallest system that proves value. https://letscode.hu/en/blog/p/not-all-tech-debt-is-equal https://letscode.hu/en/blog/p/not-all-tech-debt-is-equal Tue, 31 Mar 2026 00:00:00 GMT Treating all tech debt the same creates false progress. Strategic, accidental, and architectural debt need different decisions and different responses. https://letscode.hu/en/blog/p/ci-cd-pipeline-slower-than-your-developers https://letscode.hu/en/blog/p/ci-cd-pipeline-slower-than-your-developers Fri, 20 Mar 2026 00:00:00 GMT Developers ship fast; production does not. The bottleneck is not people—it is your pipeline, process, and architecture of delivery. https://letscode.hu/en/blog/p/adrs-are-useful-only-if-they-shape-the-decision https://letscode.hu/en/blog/p/adrs-are-useful-only-if-they-shape-the-decision Sun, 08 Mar 2026 00:00:00 GMT ADRs fail when they become passive documentation. Their real value is the thinking and discussion before the decision is finalized. https://letscode.hu/en/blog/p/your-scaling-problem-might-be-architecture-not-kubernetes https://letscode.hu/en/blog/p/your-scaling-problem-might-be-architecture-not-kubernetes Wed, 04 Mar 2026 00:00:00 GMT Before moving to Kubernetes, verify what actually breaks first. Many scaling problems come from architecture and unclear boundaries, not missing clusters. https://letscode.hu/en/blog/p/context-engineering-in-practice https://letscode.hu/en/blog/p/context-engineering-in-practice Wed, 03 Dec 2025 00:00:00 GMT Most failing agents aren't held back by a weak model - they're drowning in tool results that quietly re-inject the same 2,500-word document into every turn. The fix isn't a bigger context window; it's giving the model exactly what it needs for the current step, and nothing else. https://letscode.hu/en/blog/p/structured-output-without-asking-nicely https://letscode.hu/en/blog/p/structured-output-without-asking-nicely Tue, 21 Oct 2025 00:00:00 GMT Asking an LLM nicely for a specific output shape works until it doesn't - a few calls out of a hundred drift out of format and break a parser. Constraining the token generation itself, through tool calling or schema enforcement, is what actually makes the shape non-negotiable.