OpenClaw Adds GPT-6.1 Sol and Fixes Missing Replies Between AI Agents
Version 2026.9.8 also repairs interrupted updates and startup failures. Some custom workflows need explicit follow-up calls, and older running updaters cannot inherit the fixes.
Loading page…
Version 2026.9.8 also repairs interrupted updates and startup failures. Some custom workflows need explicit follow-up calls, and older running updaters cannot inherit the fixes.
Listen to this story
OpenClaw’s v2026.9.8 makes delegated agent work more reliable by returning results to the original requester, including later delivery when work is not immediately complete, and adds GPT-6.1 Sol for users whose OpenAI accounts already have access. The release changes workflow assumptions: automatic agent-to-agent follow-up is gone, internal sessions can no longer finish silently, and accepted work can survive connection-policy reloads while access remains valid. Teams should update custom workflows to make explicit follow-up or message calls; access revocation still cancels work, and some updater fixes apply only after the updated updater is running.
Custom workflows needing continued agent exchanges or channel delivery must make explicit follow-up or message-tool calls; REPLY_SKIP and ANNOUNCE_SKIP no longer suppress returned text.
Internal subagent and Control UI sessions must return a result or keep working; NO_REPLY remains configurable for external group conversations.
On Windows, package replacement retries temporary file-lock failures up to 16 times, with bounded waits totaling 57.75 seconds, before preserving the installed package.
An AI agent asking another agent for help should get the answer back, not lose it between sessions. OpenClaw’s v2026.9.8 release repairs that handoff and adds GPT-6.1 Sol through its OpenAI provider for accounts that already have access, according to the release notes.
When one agent delegates work, OpenClaw now returns the result to the original requester once. A finished answer can arrive immediately; otherwise, it is delivered later. Delivery uses the original requester’s identity and checks that the requester is still the same active instance.
There is a compatibility change alongside that repair. OpenClaw removes automatic back-and-forth between agents and automatic target-announcement turns. Custom workflows that need further conversation or channel delivery must now use explicit follow-up or message-tool calls. The retired loop’s REPLY_SKIP and ANNOUNCE_SKIP controls no longer suppress returned text.
The release also separates a connection change from a loss of permission. Work OpenClaw has already accepted can finish through connection-policy reloads if the user’s access remains valid. That includes accepted runs and queued inputs that previously could be cancelled simply because their originating connection retired.
Clients must reconnect before sending new requests. Revoking access still permanently cancels affected work, while changes to the authentication mode or listener setup require a restart. Rotating a local proxy fallback password cancels password-authenticated work without cancelling unrelated proxy-authenticated work.
On Windows, package replacement now retries temporary file-lock failures up to 16 times, with bounded waits totaling 57.75 seconds. If those retries run out, it identifies the affected paths and preserves the installed package. On macOS, updates recognize equivalent installation paths reached through aliases, including while npm has temporarily removed the old package directory.
Those repairs apply only once the updated updater is running. An older updater cannot acquire them mid-upgrade, including the documented v2026.9.6 first-hop limitation. OpenClaw’s Windows and macOS updates guidance also cautions that the changes do not guarantee every upgrade will succeed.
Loading discussion...
Join the conversation
Explain when automatic follow-ups would help or get in the way.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.