Addy Osmani

<start of post>

The "Context Window" is the new "Memory Management" - and most developers are still leaking tokens.

My latest deep-dive on architectural efficiency: https://lnkd.in/gX9zR2p ✍

In the early days of C, you had to be obsessed with every byte. If you didn't manage your pointers, your app crashed. Today, we have garbage collection for code, but we don't have it for context. We are shoving massive, unoptimized prompts into LLMs and wondering why the "reasoning" feels sluggish or the costs are skyrocketing.

We are entering the era of Context Engineering.

The quality of an agent's output is inversely proportional to the noise in its prompt. When you provide a 100k token context window filled with irrelevant Slack logs and outdated documentation, you aren't giving the model "more to work with." You are diluting the signal.

🎯 Dynamic Pruning: Only injecting the specific module definitions relevant to the current ticket.

🧠 Semantic Routing: Using a smaller model to decide which "knowledge chunks" the larger model actually needs to see.

⚡ State Compression: Summarizing previous agent turns rather than re-sending the entire history.

The mental model for 2026? You aren't just a prompt engineer. You are a context architect. You are building the filters that ensure the AI only sees the "truth" of the codebase, not the "noise" of the repository.

I've been experimenting with this on our internal tools at Google, and the results are clear: smaller, high-precision contexts outperform "infinite" windows every single time. Verification becomes easier, and the "hallucination" rate drops to near zero.

How is your team handling context bloat? Are you still using "select all," or are you building smarter retrieval layers?

#ai #programming #softwareengineering

<end of post>