← All writing

Category

AI Engineering

  1. Context Pruning: Shrink Data Before the Model

    Context pruning filters, pages and sums API data in a sandbox before it reaches the model. Cloudflare Dynamic Workers make that sandbox cheap and quick.

  2. Agent Skills: Load Only the Instructions You Need

    Agent Skills package instructions as folders with a SKILL.md file. Agents read a short description at startup and load the full skill only when needed.

  3. Why Agent Infrastructure Is Moving to Rust

    Codex CLI and the GitHub Copilot runtime both moved from TypeScript to Rust. What changed, the real 292 ms benchmark, and why density beats bragging rights.

  4. Code Mode for MCP: Two Tools Instead of 2,594

    Code Mode lets agents write code against an API instead of loading every MCP tool. Cloudflare covers 2,500+ endpoints with two tools and 1,000 tokens.

  5. Why Agent Sandboxes Are Moving to MicroVMs

    Agent sandboxes are moving to Firecracker microVMs. How E2B forks a live session, what the two AgentBox projects do, and why your laptop is no sandbox.

  6. Always-On Agents Mostly Sleep, and That Works

    Always-on agents are mostly asleep. Here is how an event inbox, timers, memory, permissions and an outbox let them wake, work in short runs and rest.

  7. Decision Models: The Smart If for AI Agents

    Decision models like TypeSafe's Jev and Convai's Laya return typed choices and probabilities instead of text, making agent routing fast and boring.