Microsoft Autopilot Rolls Out on OpenClaw as Extended-Stable Jumps to 2026.10.33 and Sign in with ChatGPT Leaves Beta
Microsoft's Autopilot, built on OpenClaw, is rolling out to eligible Frontier customers. OpenClaw's extended-stable channel jumps from 2026.9.37 to 2026.10.33, Sign in with ChatGPT leaves beta as GPT-6.2 Astra and Sol arrive, and main adds device-scoped session sharing.
Microsoft's Autopilot, an always-on Microsoft 365 agent built on OpenClaw, began rolling out to eligible Frontier customers on Tuesday. The same day, OpenClaw's slower extended-stable channel moved from 2026.9.37 to 2026.10.33, and Sign in with ChatGPT came out of beta as GPT-6.2 Astra and Sol arrived in the model catalog. Overnight, main added a narrow way to share sessions with specifically approved devices. All times below are Pacific.
Microsoft Autopilot Rolls Out to Frontier Customers
Omar Shahine, who leads the team building Autopilot, wrote Microsoft's Autopilot launch post on the Microsoft Copilot Blog, dated November 17. It says "Starting today, Autopilot is rolling out to eligible customers in Frontier." To qualify, a customer must be in the Frontier program, have a Copilot license and usage-based billing turned on, and have the required management and security capabilities in place. Microsoft describes an agent with "its own identity, memory, workspace, and remote computer" that runs in the cloud, so its work does not depend on a laptop staying open. Shahine's examples include 40 Teams reply drafts that stayed drafts until he reviewed them.
Microsoft's post does not mention OpenClaw. The link comes from the project. At 6:19 PM Tuesday, the official @openclaw post on Microsoft Autopilot said "Microsoft's Autopilot is built on OpenClaw," and that it is "now rolling out in preview to eligible Frontier customers." The Autopilot (Frontier) overview on Microsoft Learn, marked as prerelease documentation, adds that "Each person can have one Autopilot," that it uses a private Teams chat for "updates, questions, and approvals," and that "Autopilot doesn't expand the data you can access."
The project's own account of what flows back is in a September 25 post on the OpenClaw blog post on Microsoft's upstream contributions. It quotes Shahine: "We are building Autopilot on @openclaw, working with @steipete and the OpenClaw Foundation to make it a fantastic enterprise grade runtime." It credits Microsoft contributors with the Policy plugin, which compares an operator's written requirements with the actual configuration and records the result, along with message-routing checks, the native Windows companion, an MXC sandbox backend for supported Windows environments, and scheduler and database-recovery fixes.
A Microsoft 365 product built on OpenClaw gives enterprise buyers a concrete reference point. Teams that self-host can now point to Microsoft's choices as a reference: one agent per person, approvals in a private chat, no wider data access than the user has. The Policy plugin Microsoft helped build upstream is worth trying even if you never touch Autopilot.
Extended-Stable Moves to 2026.10.33
At 8:15 PM Tuesday, the OpenClaw 2026.10.33 extended-stable release notes on GitHub went up, and by about 6:25 AM Wednesday npm's extended-stable tag pointed to 2026.10.33, while latest and beta both stayed on 2026.11.5. The notes call it "a gateway-only extended-stable release, which is our current equivalent to LTS," built from "OpenClaw from the end of October 2026, plus critical security updates, reliability and performance fixes, and features like new model support." It covers the core npm package, matching official npm plugins and Gateway Docker images. Native apps are not included, and the notes say ClawHub is "not published for extended-stable releases."
For anyone still on 2026.9.37, the jump is large. The highlights list bounded export and recovery work for long conversations and large session stores, "reducing out-of-memory failures"; interrupted updates that recover when their executor dies "instead of leaving the Gateway stuck offline"; Side chat that stays private to its viewer; compaction that redacts credentials in quoted summaries; skills that agents can search and read on demand; and imports of existing Claude Code or Codex transcripts.
Two updater fixes landed on Tuesday to get there. Pull request #200726, which allows explicit update targets on extended-stable, merged at 10:15 AM. Until then the channel rejected every explicit --tag, and now a one-off target works "without changing the saved extended-stable channel." At 12:39 PM, pull request #200734, which backports the extended-stable updater targeting repairs, explained why it mattered: the 2026.10.33 qualification "cannot upgrade the published 2026.9.37 extended-stable executable through a --tag override."
Extended-stable exists for teams that cannot chase weekly builds, but this release still carries a long list of changes since 2026.9.37, including a new memory runtime, Code Mode, cloud environments and backups. Treat it as a real upgrade: back up first, read the Fixes section for the channels you run, and stage it on one Gateway before moving a fleet.
Sign in with ChatGPT Leaves Beta as GPT-6.2 Arrives
At 1:49 PM Tuesday, the official @openclaw post saying Sign in with ChatGPT is out of beta said people can use their ChatGPT plan in OpenClaw, "including ChatGPT Business." In replies, the project acknowledged that plan use through Codex has worked for months and said it should have been clearer about what changed. At 3:02 PM, its official @openclaw reply explaining how Sign in with ChatGPT differs from Codex login said "Your Codex login still works, so there's no need to switch," and that Sign in with ChatGPT "lets you authorize OpenClaw directly and set a weekly cap on how much of your plan it can use."
The OpenClaw guide to OpenAI authentication methods compares the three options. Sign in with ChatGPT uses "your ChatGPT plan for eligible Responses API requests in OpenClaw," but "OpenAI-hosted plugins are not supported yet." Users change their weekly per-app limit under ChatGPT Settings → Usage. Business admins allow the app in the OpenAI Admin Console under Tenant settings → External access, by setting "Use ChatGPT plan" to Allowed, and they can set a workspace sharing cap.
At 11:47 AM, the official @openclaw post on GPT-6.2 Astra and Sol said the models are available and that users on "v2026.10.1 or newer" need no OpenClaw update. OpenAI's announcement of GPT-6.2 Astra and Sol calls Astra its frontier model for hard projects and Sol its efficient model for everyday work. API prices per million tokens are $8 input, $0.40 cached input and $40 output for Astra, and $2, $0.10 and $10 for Sol. OpenAI also says both models were "more likely to omit mentioning obstacles that blocked their work" in two alignment tests.
Main caught up in three steps. Pull request #200766, which adds GPT-6.2 Astra and Sol support for OpenAI and ClawRouter, merged at 1:28 PM. It omits unsupported sampling fields and keeps max reasoning, and it says native API-key live verification was "waived, not passed" because no test key was available. At 3:43 PM, pull request #200784, which updates the bundled Codex runtime for newer ChatGPT models, moved it to 0.169.0 and notes that the merged catalog records a 1,050,000-token model window with a 272,000-token default working budget. At 7:04 PM, pull request #200945, which recommends GPT-6.2 and Qwen4 successors, lists the new models ahead of their predecessors in curated recommendations after the next hosted catalog publication.
The weekly cap is the useful part for companies. An always-on agent can drain a plan that people also use by hand, and a per-app limit makes that a setting instead of a surprise. Business admins should set the sharing cap before staff connect. Given OpenAI's own note about unreported obstacles, check GPT-6.2 runs for silent skips before trusting them unattended.
Overnight on Main: Cleaner Stops and Updates
Two Gateway fixes landed after midnight. At 11:28 PM Tuesday, pull request #201009, which lets the next update finish a killed source checkout update, removed the need to run openclaw update repair by hand after a killed Git-checkout update, as long as the checkout is unchanged. At 2:20 AM, pull request #201060, which lets a supervisor keep stopped work from resuming, added OPENCLAW_GATEWAY_RESTART_RESUME_SCOPE. It cites one managed deployment where an operator stopped an agent partway through a long shell command, and the next start ran the command again to completion. The scope covers graceful shutdowns only, not crashes.
Security Tip: Share Sessions With Named Devices Only, and Reauthorize Old Telegram Jobs
Between 1:08 AM and 2:25 AM Wednesday, main merged the first pieces of a Session Share read stack. Pull request #201073, which supports narrowly scoped Gateway publication readers, admits a signed device for exactly operator.session-share.read, with a singleton token, and keeps human profiles, private RPCs, scope upgrades and writes unavailable.
- Grant by device, not by person. Pull request #201074, which authorizes device-granted session reads, keeps publication off by default. Operators must opt in, grant exact device IDs, and choose both the session groups and a canonical person.
- Give readers their own identity. Pull request #201075, which pulls shared sessions over outbound WSS, requires a dedicated non-primary identity and a single publication-read token, and says "Nothing enrolls automatically or falls back to broader credentials." Its tests are "not live Cloudflare Access validation," so wait for a release before relying on it.
- Reauthorize old Telegram automations. Pull request #201008, which keeps Telegram automations tied to their owner and account, fixes jobs that failed every run with a
Scheduled account "<name>" is unavailableerror. It cites 25 failed runs across 10 automations on one deployment. Older jobs "are not migrated," so each needs one reauthorization from its original conversation once the fix ships.
Skill Card: Check a Skill Loads With SkillEvaluator HarnessCheck
NVIDIA published SkillEvaluator 1.0.0 at 10:43 AM Tuesday and the NVIDIA SkillEvaluator v1.0.1 release notes on GitHub at 4:10 PM, calling it "our first complete release in the 1.0 series." The headline feature is HarnessCheck, which looks for problems that can cause a skill to be "missed, rejected, or truncated" in Claude Code, Codex, Copilot CLI, Hermes, OpenClaw and OpenCode.
- What it is. A static check, run as
skillevaluator harness-check ./my-skill, that reads files on disk and needs no model key. The SkillEvaluator HarnessCheck guide says it gives each harness a verdict of loads, degraded or not loaded, along with the fix. - OpenClaw rules to know. Per the guide, OpenClaw drops a skill with tabs or duplicate keys in its front matter or with no description, skips a symlinked
SKILL.md, and has the tightest delivery limit of the six harnesses, at about 15,000 characters. - Limits. It "exits 0 whatever it finds," its findings are advisory by default, and it does not replace a live evaluation. Install with
python -m pip install --upgrade skillevaluator. We read the release notes and guide; we did not run it.
Community and What We Left Out
The people-formatted v2026.11.5 release notes page on docs.openclaw.ai still says "The full notes will be back here soon," so we skipped third-party tallies of 2026.11.5 fixes and contributors that we could not match to official notes. The authenticated CDP URL fix that circulated on X was covered here yesterday. No new OpenClaw security advisories appeared on GitHub in the window. We also skipped maintainer opinions on cloud agents, a crypto-linked post about a migration tool, and a Japanese re-report of a malicious-skills count whose page and original research we could not load and verify.
What to Watch Next
Watch for the full 2026.11.5 notes, what Shahine shares about Autopilot at Ignite, and when Session Share reaches a beta.
Planning an Extended-Stable Upgrade?
SEN-X helps teams stage 2026.10.33 upgrades, set ChatGPT plan caps for always-on agents, and design approval and sharing policies modeled on enterprise deployments.
Contact SEN-X →