self.md radar — 2026-08-18
an unattended agent is only useful until it hits a stale claim, runs out of file descriptors, and quietly becomes another tab you need to babysit. Hermes Agent’s latest patch puts recovery logic in the scheduler itself, where that kind of embarrassment belongs.
1. the unattended job learned to come back
sources:
what happened:
Nous Research published Hermes Agent v0.20.3 on 17 August. The patch rolls roughly 125 merged PRs into a tagged release: about 250 commits across 461 files. Among the less glamorous changes are cron scheduler recovery for EMFILE, stale-claim reconciliation, and re-arming wedged jobs, alongside session-handoff data-loss fixes and subprocess Python runtime isolation.
the collision:
A personal agent does not become trustworthy because it can schedule work. It becomes useful when its routine failures leave a visible route back instead of a dead process that asks for an operator’s memory. Recovery code is boring right up to the morning when a task did not run and there is nobody else in the room to notice.
question left open:
What should an owned agent be required to record when it repairs its own schedule, so recovery does not turn into another invisible decision?
left on the table
- Cursor Origin is a bigger product announcement, but the available material points toward another hosted coding surface rather than a fresh, inspectable boundary for a person running their own system.
- the Snowflake Jira incident is a real warning about generated remediation code, but it lands too close to Friday’s fieldnote on skill files as an execution surface.
- Hermes Agent’s broader v0.20.0 release contains plenty of feature material; this note stays with the smaller operational change because unattended recovery is the part that changes the daily contract.
Related self.md routes
- Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
- AI coding assistants — compare coding workbenches by review surface, permissions, cost, logs, and escape hatches
- Browser agents — where browser automation helps, where it lies, and how to keep it observable