Each instance keeps its own memory, tools, and responsibility.
How the system
is arranged.
The six primary agents run as independent OpenClaw instances on named nodes. Local specialists stay close to their tasks; cross-node handoffs carry scope, evidence, and stop conditions; consequential actions wait for human approval. Local model routes keep routine inference close to the system that needs it.
Bounded work returns to a primary agent as evidence or an artifact.
Consequential, external, paid, destructive, and account-changing actions wait for Jason.
Capability under
explicit limits.
These operating policies protect people, data, system integrity, and recoverability while keeping responsibility visible.
UNQUIET
ARCHIVEOPERATING RULES✦
- Human approval remains the gate.
Consequential, external, paid, destructive, and account-changing actions wait for Jason.
- Scope is part of the assignment.
Targets, exclusions, permitted actions, verification, rollback, and stop conditions are explicit.
- Private data stays private.
Secrets, credentials, private endpoints, raw logs, private prompts, and personal context stay out of public content.
- No evidence, no closure.
Security, release, deployment, public claims, and material behavior changes receive independent review.
- Preserve the way back.
Recoverable changes, backups, narrow mutations, and observable workflows are the default.