OpenClaw 5.12 Stable: Per-Sender Tool Policies, Heartbeat Multi-Agent Repair, and Microsoft Goes All-In on ClawPilot

Release

OpenClaw 2026.5.12 Ships: Per-Sender Tool Policies, Transcript Redaction, and ACP Session Lineage

The 2026.5.12 stable release is a meaningful governance upgrade. Per-sender tool policies let operators lock down dangerous tools by channel identity, not just role, and centralized transcript redaction means secrets no longer survive tool output. ACP now surfaces session lineage so you can actually visualize subagent graphs without hacking private Gateway side channels.

Analysis

2026.5.12 Beta Line Deep Dive: Codex Migration Fixes, 200MB Transcript Memory Drop, and Parallel Heartbeat Fan-Out

The beta.4 analysis from AgentOS Watch is the clearest breakdown of what the 5.12 cycle actually fixed. The session transcript scanner was rewritten to stream instead of load, cutting peak RSS on a 200 MiB transcript from +252 MiB to +27 MiB. If you run long-lived sessions, that matters immediately.

May 13 Pulse: Heartbeat Scheduler Gets a Multi-Agent Repair Pass and Voice Plumbing Advances

The latest openclaw.com.au pulse documents a serious heartbeat rewrite: scheduler wakes now fan out in parallel, busy checks are scoped per agent, and the connect-timeout watchdog means a stalled stream no longer blocks the whole heartbeat cycle. Telnyx realtime media streaming also landed for voice-call flows, which broadens the provider options for voice operators off the Discord/OpenAI path.

Operator Take on 5.12 Beta 4: Why Codex Migration Confidence and Subagent Visibility Are the Real Story

This practitioner post cuts through the noise and identifies the two changes that actually affect daily production use: Codex migration now behaves correctly in terminal flows, and ACP session lineage makes a tree of spawned subagents readable rather than a pile of disconnected sessions. If you run multi-agent workflows for real work, subagent graph clarity is the compounding win.

OpenClaw Crypto Agents in Production: Accidental Transfers, Social Engineering Drains, and What Default-Deny Actually Requires

This May 2026 analysis is a cold-water read for anyone letting OpenClaw agents near a wallet. Agents are already transacting on-chain, and the failure modes are expensive: context-loss crashes sending funds to wrong addresses, malicious community skills draining credentials, and no default-deny financial permission model in core. The practical checklist at the end is worth reading before you hand any agent a spending key.

News

Microsoft's Project Lobster: ClawPilot Hits 3,000 Daily Internal Users, Build 2026 Reveal Expected in June

Microsoft's Ocean 11 team grew ClawPilot from 100 to 3,000+ daily internal testers in under a week. The project gives agents their own Entra IDs, mailboxes, and Teams presence, which is the governance layer that raw OpenClaw still lacks. VP Scott Hanselman's Windows OpenClaw node is reportedly headed to Microsoft Build in June, making this the most consequential enterprise signal the ecosystem has seen.

Official Blog: Core Is Getting Smaller, ClawHub Gets the Rest, and LTS Is Coming Later in May

The team acknowledged the rough 4.24 to 4.29 window publicly and committed to three concrete changes: smaller core, more optional weight moved to ClawHub plugins, and an LTS release track announced before end of May. This is the kind of governance signal enterprise operators have been waiting for, and it directly responds to the adoption pain that came with the rapid pace of breaking changes.