Notes from the runtime
How Waveguide is built and why — the security architecture, the economics, and the thinking behind an agent you can trust with real accounts.
Missions: giving an agent a goal instead of a task
Task-based AI does what you say and stops. A mission is a goal your agent pursues for weeks — research, outreach, replies, booked meetings — inside rails that make autonomy safe.
Honest metering: multi-model routing and the 80/100/120 budget ladder
How an always-on agent stays affordable: route every job to the cheapest model that does it well, suspend the machine between wakes, and bound the worst-case bill by construction.
Approvals by construction: how T0–T3 risk tiers work
Every action your agent takes is classified into one of four risk tiers, and the dangerous ones physically require a human. Here's the taxonomy and why it's enforced in the network layer.
Why the model never holds a credential
The single design decision that makes Waveguide trustable with real accounts: OAuth tokens live in a vault, and a proxy injects them after the request leaves the model.
Subscribe via RSS.