October 13, 2026 • 9 min read
A pilot-style preflight checklist for AI agents: capability leases, deny-by-default tools, evals first, attested MCPs and skills, cost routing, vendor-multi, reproducibility, and the rest of what must be true before takeoff.
October 12, 2026 • 11 min read
Where AI authority ends: a reference architecture for retrieval-backed, human-approved workflows, with three boundaries, a state contract, and a tested approval gate that rejects a recipient swap.
October 12, 2026 • 8 min read
How SKILL.md files let agentic applications, coding harnesses, and assistants like OpenClaw load specialized knowledge only when the task needs it.
October 3, 2026 • 30 min read
Can a decision model acting as a reflex manage traffic and battery for 10 drones on a shared map? A 30-minute experiment with and without a deterministic safety backstop: 12 collisions without it, none with it.
October 3, 2026 • 8 min read
How generative AI routing differs from load balancing: KV caching, model affinity, switching budgets, and when to use an inference router, an AI gateway, or both.
October 2, 2026 • 30 min read
A map of the AI-native cloud (managed agent runtimes, serverless inference, routing, caching, knowledge bases, evaluations, GPUs) followed by a product-by-product tour of what a full AI cloud offers, with illustrative pricing.
October 1, 2026 • 6 min read
Jev from TypeSafe AI understands language at a frontier level but never generates a token. It is a non-autoregressive decision model on DigitalOcean Serverless Inference that returns typed answers with calibrated probabilities. Here is how it differs from an LLM, with working API calls for all three question types.
August 27, 2026 • 8 min read
A breakdown of pstack, a heavyweight AI agent orchestration framework, compared against a lightweight SCOPE-based large-brain-to-small-brain pipeline: two different paths to the same engineering rigor.
August 26, 2026 • 4 min read
A worked example of grading an AI agent's response against its own tool-use and formatting instructions, and why 'it produced a correct answer' isn't the same as 'it followed the rules.'
August 18, 2026 • 7 min read
Coding agents like Claude Code load a fixed startup budget of system prompt, tools, memory, and skills before you type a single word. Here's how to audit that budget and prune the biggest offender: bloated memory files.