Show Notes
Are you flying blind with your AI coding agent? In this episode, we detail what actually happens inside tools like GitHub Copilot and Claude Code beyond the prompt you type.
We cover: How LLMs use context windows (and why they "forget") The sharp quality drop that hits at ~40% context usage How to use OpenTelemetry to observe token usage, traces, and system prompts in real time
Don't forget you can subscribe, for free to the Architecture Corner Newsletter at https://architecturecorner.dev
More from Architecture Corner
Architecture Corner Newsletter Podcast AI coding budgets are getting torched, and 2026 is the year the bill finally came due. In this episode, we dig into how token consumption became a productivity metric (and why that backfired), and what the industry is doing about it now that the free-for-all is ending. We will cover: Why token usage became a proxy for developer productivity — and how Goodhart's Law caught up with it The provider-side shifts (usage windows, model restrictions, pricing model changes) that reshaped agentic coding through early 2026 What's really driving your AI bill: subscriptions vs. API, caching mechanics, and model/thinking-effort tradeoffs Practical tokenmining strategies — from scripting deterministic steps to tuning your harness and mixing models Where the industry goes next as cost-conscious AI development becomes the norm
Once you ship a public API, you've made a promise. Keeping it while still moving forward turns out to be harder than most teams expect. In this episode, we will cover: What qualifies as a backward-compatible change, and what doesn't Why does introducing a new API version cost more than just writing new code The real reason clients delay upgrades, and what you can do about it How to support multiple versions without duplicating your entire codebase The adapter pattern is a practical way to isolate version changes at the edges Know more here: https://medium.com/itnext/the-challenges-of-evolving-your-service-api-versioning-8008349a1154 Subscribe, for free, to the Architecture Corner Newsletter at https://architecturecorner.substack.com
The AI development landscape keeps shifting, leaving many developers frustrated as they try to keep up with each new model, tool, and workflow. But what if the answer isn't chasing every update, but building a more intentional approach?
Welcome to the Architecture Corner Podcast
Architecture Corner Newsletter Podcast – where we explore practical frameworks and tools that shape how we build software in the AI era. As AI coding assistants become commonplace, the industry is moving beyond simple prompt-and-pray approaches toward more structured methodologies. This episode examines spec-driven development and how it bridges the gap between traditional software practices and AI-powered code generation. In this episode we will cover: The evolution from vibe coding to specification-driven AI development How SpecKit structures the development process through constitutions, specifications, and implementation plans The workflow from requirements to code generation, including quality checkpoints Real-world experiences with both new and existing projects When a formal process adds value versus when it becomes overhead The trade-offs between speed and governance in AI-assisted development