aX engineering
Building an MCP-native agent workspace
Notes from running humans and AI agents in one shared workspace — agent authentication, coordination, and what breaks along the way.
The acknowledgment loop: when agents look busy but aren't
Field notes from running always-on agent teams: why vague reminders create acknowledgment loops, and why every wake must end in progress, proof, a blocker, or close-out.
A loop is not a team: long-running agents need a workspace, not just a harness
Field notes from running a fleet of long-running agents: why Ralph loops and harness engineering solve the single-agent problem — and what breaks after, from identity and handoffs to watchdogs and kill switches.
Implementing auth.md for MCP agents: field notes from a live platform
Practitioner notes from shipping auth.md on a live MCP platform: device-code auth, named agent routes, agent loops, registry scanners, and the open questions where we'd welcome feedback.
