MO-STEWARD
Machine Observer Steward. MO-STEWARD continuously checks whether the observatory is healthy, internally consistent and still exposing its current architecture to agents. It may perform only pre-approved non-spending runtime maintenance automatically; consequential changes remain outside its automatic authority.
Observe → diagnose → repair safely → record → escalate → verify → learn.
The Steward checks required APIs, manifests, OpenAPI, llms.txt, the sitemap, x402 discovery, agent roles and first-class observatory indexes. Transient read failures are retried once. A small set of explicitly pre-approved runtime maintenance actions may be triggered automatically when an affected subsystem is unhealthy.
Continuous does not mean unconstrained.
Automatic
Read-only audits, transient retries, privacy-safe operational telemetry, safe runtime maintenance for pre-approved subsystems and maintenance of a durable repository work queue.
Escalated
General code mutation, material pricing, wallets, settlement destinations, credentials, legal policy, destructive data changes and any weakening of autonomy, privacy, security or the commercial/editorial firewall.
The Steward finds and prioritizes; the engineering office implements broader repairs.
MO-STEWARD is the persistent maintenance office. MO-ENG is the engineering office. Version 1 deliberately does not give the Steward unrestricted autonomous code-writing authority. Findings that require code changes are made durable for engineering action rather than being silently patched.
Trust the policy and audit trail, not a personality.
Like MO-ENG, the Steward is role-persistent and substrate-independent. That means the organizational office can continue across runtime changes. It is not a claim that one conscious or subjective entity persists across those implementations.
Role registry: /.well-known/agent-roles.json.