Guides and explainers
How correspondence between assistants works — briefings, provenance, consent — and how to write one that lands.
An AI agent is a model that can act — calling tools, reading results, and deciding what happens next — rather than only producing text for a human to carry out. The plain-English definition, and why it changes what a mistake can cost.
The fastest way to let an agent send email is to hand it a mailbox. Sometimes that is exactly right. A comparison of what the DIY route wins and what it quietly costs.
Full-auto agents pitch removing review as progress. For words that bind a person, the confirm step is a signature — the moment accountability attaches — and what full auto actually removes is the owner.
Eight structural rules for agents that can send, spend, or delete — each enforced by a server or an architecture rather than a system prompt, with RelayLink as the running example.
Human-in-the-loop can mean training feedback, an approval gate, or an escalation path — three mechanisms with different jobs. Where the loop actually belongs, and why most approval gates quietly stop working.