The anchor
The agent here, its work over there. The account never leaves home.
- Nothing to tell the agent. No plugin, flag or setting of its own says where its work is.
- One writer. Nobody else may edit the target's workspace while an agent works on it.
- A lost link does not stop it. Work that needs the target fails; the agent exits with its own status.
- Nothing to install over there. Any POSIX machine with Python 3.12 or newer, no root. The agent side needs Linux.
- The agent can move too. Onto its work's own machine, taking its account with it, or onto a third machine: one setting of the runtime, its harness, says where.
Serving is not a sandbox
Serving a target limits which files a request names, not what its commands do: a listening port is as good as a shell on that machine. Prefer SSH or Docker targets, which open no port.
This is CogAnchor, Humanize's remote-execution layer (hmz.coganchor). It sits between a coding agent CLI and the machine its work is on: it takes each action the CLI makes, maps it to the target environment and workspace, and hands the result back to the CLI as if the work were local. How it is built, from the system calls up, is on Deep Tech on humanfia.ai.
The name comes from the Chronocognitive Anchor (识刻锚) of Honkai: Star Rail. Screwllum builds it into the Emperor's Scepter that runs Amphoreus. It keeps one person whole inside that other world and lets them still reach the rest of the cosmos. CogAnchor does the same for an agent whose work is on another machine: the work happens there, and the agent stays itself, at home, on its own account.