The short answer

A network partition leaves peers with incomplete or conflicting evidence. Agent Mesh preserves the context needed for reconciliation; the application chooses the permitted operating policy. Connectivity loss must not create new permission for irreversible effects.

A partition changes what a peer can know

A peer may remain healthy and execute local code while losing contact with the evidence needed to justify its next action. The problem is not necessarily a crashed process; it can be an incomplete view of assignment or membership.

The Mesh failure reference treats stale, equivocal and uncovered evidence as unresolved. Applications can pause, narrow eligibility or recover after reconciliation. Avoid equating successful local computation with permission to continue every operation.

Choose a partition policy before the incident

The Collective Runtime partition reference names four policies. Strict consistency requires quorum for sensitive operations. Bounded degraded operation limits work to reversible or bounded effects under risk budgets. Availability-preferred operation stays within pre-authorized limits. Reconciliation-required operation blocks new effects until causal branches converge.

These names are not permission to improvise after disconnection. Configure what the deployment allows beforehand, including which work must stop and which evidence is necessary for resumption.

Separate useful local work from irreversible action

Consider an illustrative peer that can continue drafting an internal analysis during a partition. That may be useful reversible work under its configured policy. Sending a binding instruction to another system is a different operation with a different consequence.

The partition state machine explicitly preserves a restriction on irreversible effects without quorum or reconciliation. If your application cannot classify a proposed action within the allowed boundary, do not silently relabel it as harmless because local execution remains available.

Reconnection starts reconciliation

The documented lifecycle includes connected, degraded, partitioned and reconciling states before convergence or safe stop. A restored socket therefore does not settle which branch is current.

Use causal synchronization and required membership context to reconcile the affected work. Then revalidate the intended assignment and effect. A permit or local snapshot from before the partition may no longer be sufficient for the current transition.

Test the allowed and denied operations

  • Record the selected policy and its pre-authorized limits.
  • Partition the peer while a specific operation is pending.
  • Observe which work pauses and which bounded work proceeds.
  • Attempt an irreversible effect without the required evidence.
  • Reconnect, reconcile and inspect the reason for resumption or safe stop.

Retain both denied and completed cases. A report that only counts local progress can hide whether the system preserved its intended effect boundary during the incident.

Sources and further reading

Documentation reviewed . Consult the linked documentation for current implementation details.