Field notes on agents, MCP, and shipping
First-person notes on new models, dev approaches, and the infrastructure we ship in production. Less guide, more logbook. Written for engineers, not buyers.
RSS feedThe 70% tax on agentic loops (and why MCP makes it worse)
Roughly 70% of the tokens an agent sends are operational overhead - system prompts, tool definitions, retrieved context, accumulated history. MCP servers make the tool-definition tax bigger. Here is the math and the fixes.
OpenShell vs Docker Sandboxes: governance vs isolation
Both run on your laptop and both keep an agent from wrecking your machine, but they optimize for different things. OpenShell is a policy engine. Docker Sandboxes is a microVM. Here is when to pick which.
E2B vs Crabbox: a sandbox runtime vs an orchestrator
They look like competitors. They are not. E2B is a managed sandbox runtime (Firecracker microVMs). Crabbox is a control plane that brokers runners from many providers, including E2B. Here is how to think about them.
One guardrail file, every coding agent
Every team I have worked with ends up with three coding agents and zero shared rules. The fix is one guardrail file the agents read before they touch anything.
Why I built an autonomous task queue for Claude Code
Claude Code is great at a single task. It is bad at remembering what comes next. Sugar is the layer that fixes that - an open-source task queue that lets coding agents work in the background.
Welcome to Eng Labs
What this section is, how it differs from the blog, and what to expect - short field notes on new models, MCP, agent architecture, and shipping production AI.