https://mkabdelrahman.github.io/tags/agents/Recent content in Agents on Mohamed AbdelrahmanHugoenMon, 04 May 2026 00:00:00 +0000https://mkabdelrahman.github.io/posts/lernen-ai-tutor-watching-keystrokes/Mon, 04 May 2026 00:00:00 +0000https://mkabdelrahman.github.io/posts/lernen-ai-tutor-watching-keystrokes/<h2 id="a-flashcard-app-but-the-ai-is-in-the-loop"> A flashcard app, but the AI is in the loop <a class="heading-link" href="#a-flashcard-app-but-the-ai-is-in-the-loop"> <i class="fa-solid fa-link" aria-hidden="true" title="Link to heading"></i> <span class="sr-only">Link to heading</span> </a> </h2> <p>Most &ldquo;AI study buddies&rdquo; are chatbots glued to the side of a flashcard app. You study, then you ask a question, then the model answers from cold context. The model has no idea you just got the same word wrong twice and replayed the audio four times.</p> <p>I wanted to flip that. The model should see every action — reveals, replays, grades, notes — and react inside the study loop, not after it.</p>https://mkabdelrahman.github.io/posts/agent-context-compaction-strategies/Thu, 30 Apr 2026 00:00:00 +0000https://mkabdelrahman.github.io/posts/agent-context-compaction-strategies/<h2 id="the-problem"> The problem <a class="heading-link" href="#the-problem"> <i class="fa-solid fa-link" aria-hidden="true" title="Link to heading"></i> <span class="sr-only">Link to heading</span> </a> </h2> <p>A long-running agent accumulates messages over time: user prompts, model responses, reasoning traces, tool calls, tool outputs. The model&rsquo;s context window is finite — from a hundred thousand tokens on the smaller end to a million or more on the largest current frontier models. Once the conversation no longer fits, continuing without intervention triggers a hard error or silent truncation by the provider.</p>