An agent with a 1M-token context still meets you as a total stranger tomorrow — the context window is RAM, wiped the moment the session ends. Real memory is an engineering system you build around the model: what to save, when to update it, and what to forget. This is the taxonomy, the storage trade-offs, and the write policy that separates a gimmick from a colleague.
Your browser now ships a language model. Chrome and friends expose on-device AI through a family of built-in APIs backed by a small local model, so inference can run entirely on the machine — private, free, offline, and low-latency. It's not here to replace GPT; it's a new tier you route the easy jobs to.
An AI SDK doesn't make the model smarter — it deletes the streaming parser, the tool-calling loop, the JSON-validation prayer, and the per-provider glue you'd otherwise hand-roll against a raw HTTP endpoint. Here's exactly what you get over `fetch`, using the Vercel AI SDK, and where the abstraction leaks.
A microservice fails loudly — a 500, a stack trace, a pager going off. An LLM app fails politely: the answer just got worse, the cost quietly tripled, or the agent looped fourteen times before answering, and nothing errored. This is how you point OpenTelemetry at that problem, so 'the model felt off' turns into 'the third tool call took six seconds and triggered two extra model round-trips.'
Two protocol standards keep getting confused in architecture meetings, and they solve perpendicular problems. MCP standardizes how one agent talks down to its tools; A2A standardizes how agents talk sideways to each other. Here's the anatomy of both, the mental model that makes the difference stick, and an honest case for why most systems still shouldn't go multi-agent.
A working list of the Claude Code plugins I keep installed across machines — what each one does, where it pays for itself, and which ones I uninstalled within a week. Not a feature tour; a working setup.