Topic, page 2
Page 2 of articles on Architecture.
Transcription accuracy on clean audio is a solved problem. Production audio is noisy, overlapping and full of the terms your domain depends on.
Once you have sent a 200 and started writing tokens, there is no status code left for the failure that happens halfway through the answer.
Asking politely for JSON works most of the time, and most of the time is a poor guarantee for something a parser sits behind in production.
A system prompt is the one piece of a feature everybody edits and nobody owns. Structure decides whether it stays coherent or becomes sediment.
Cost, latency and the point at which text gets cut off are all counted in tokens. Reasoning about the system in words will mislead you in all three.
When an agent misuses a tool, the instinct is to rewrite the prompt. Usually it's the tool's own signature that made the mistake available in the first place.
Request logs tell you a call succeeded. For a model-driven system the useful record is the whole chain, because the mistake is rarely in the last step.
A model that describes an image accurately is not yet a system. Frame rate, cost per frame and what happens on a wrong answer decide the architecture.
Most things called agents are workflows with a model bolted on. The line is whether the system picks its own next step, and everything difficult follows.
MCP gets described as USB-C for AI tools. The useful question is narrower: which parts of tool integration stop being your problem, and which never were.
Running a model locally is free in the sense that a server is free. The bill is idle capacity, and the arithmetic decides this more than principle does.
An MCP server turns an internal API into something a model calls on its own initiative. The threat model changes, and the existing auth boundary rarely fits.
Whether it needs designing from scratch or rescuing from its own success, tell me what you’re building and I’ll tell you how I’d architect it.
Open to remote and hybrid work worldwide