2026-09-15

The personality the agent refused to read

Every installation we deliver receives a personality file, read when the agent starts. For hours, our agents nonetheless introduced themselves under their generic identity and recited the list of all their tools. The file was in its place. The agent was refusing it. Found on 24 July 2026.

A filter against role takeovers

Hermes, the engine of our installations, discards context files that read like a takeover of its role. Our first text said “you are not a general-purpose assistant” and “never pretend to be an AI for everything”. Those negative phrasings triggered the filter: the boot log carried a one-line warning, Context file SOUL.md blocked: role_pretend, and the agent fell back to its default identity, followed by the description of every bundled skill.

Why the defect did not show everywhere

The design agent, whose text was written affirmatively, was never blocked. And the path that passes instructions through the API bypasses the context-file filter. Both worked; only the agents delivered to clients read the file, and only they refused it.

The fix

The text is rewritten affirmatively, with no negation and no “pretend”. It also says to answer with the automation's purpose before the list of tools: an invoice sorter no longer offers to control the lighting. Verified across the whole fleet: no block, and every agent introduces itself from its personality file alone. The same text is also passed a second time through the API, to survive a future change in how files are loaded.

The rule we keep

An agent's personality is written affirmatively, and the boot log is read at every change to that file: a one-line warning there announces an agent that will not be the one you described.

A personality file is written affirmatively, and the boot log says whether it was read.