OpenClaw merged a major ownership change for multi-agent installations. PR #114388, titled "refactor(agents): make multi-agent ownership explicit (H2-1 core)," retires the old ambient default-agent marker and replaces it with explicit ownership rules across sessions, channels, automations, task runtimes, authorization paths, and Gateway routing.
The problem was not that single-agent installs were confusing. They can still pick the sole available agent naturally. The risk appears when a roster contains more than one agent and OpenClaw has to decide who owns a bare session key, channel binding, cron row, fixed store, or task. A persisted default: true marker could become stale after roster changes or restart, which meant the system might infer an agent where it should have required durable ownership evidence.
Why It Matters
Multi-agent setups are becoming more common as users separate personal agents, work agents, local runners, Codex-style coding agents, channel-specific bots, and experimental workers. In that world, "default" is not just a convenience. It is a security and data-boundary decision.
The PR describes cases where the old model could select the wrong agent, lose a legacy owner, or attribute colliding bare session keys such as global to a requester instead of the durable owner. That can affect more than UI routing. It touches memory, transcript lookup, task ownership, broadcasts, active runs, channel bindings, and authorization inheritance.
What Changed
OpenClaw now uses sole-agent-or-explicit selection instead of a stored ambient default marker. A write-once agents.ownership: "explicit" stamp records intentional roster ownership. Legacy ownership is retained as migration evidence and compatibility input, not as an open-ended default.
Before retiring the marker, OpenClaw materializes compatibility ownership into the surfaces that need durable identity:
- Channel bindings and auth inheritance
- Heartbeat, system-agent, and Talk settings
- Configured fixed session stores
- Session aliases, IDs, and transcript lookups
- Task runtimes and active-run ownership
- Migration, recovery, and Doctor repair paths
The fail-closed behavior is the important part. If a multi-agent install cannot resolve a conflicting or ownerless request safely, OpenClaw now returns an actionable explicit-selection response instead of silently adopting a default.
User Impact
For single-agent installs, the expected workflow remains intact: the sole agent can still be selected without extra configuration.
For multi-agent installs, operators should see clearer behavior when a request needs ownership. Existing installations get migration help that preserves legacy owners where OpenClaw can prove them, including bare fixed-store sessions. Conflicting explicit agents and ownerless authorization requests fail closed.
Doctor also gains repairs for marker retirement, fixed-store ownership, heartbeat cadence, and startup-owned cron rows. Remote clients can consume the new additive Gateway metadata around ownership and selectionRequired without breaking older protocol consumers.
Verification
The PR shipped with full pull-request CI, changed-surface gates, full type and build lanes, and 952 touched tests across 11 Vitest shards. It also includes additional task-owner repair tests, Plugin SDK contract reconciliation, CodeQL/security checks, and explicit paired-agent regressions for bare global session ownership.
The Bottom Line
PR #114388 is a platform-level hardening change. It makes OpenClaw treat agent ownership as durable identity, not ambient preference. That is exactly the direction multi-agent systems need as they move from clever local workflows into long-running, channel-connected infrastructure.