What's New in Bots: Nov 3, 2026

OpenClaw shipped a big release with an Advisor that can catch an agent going off track, Claude got faster long conversations, and OpenAI built a test for everyday tasks, so here is what is new in bots.

OpenClaw: an Advisor for your agents

OpenClaw released v2026.10.10 overnight, a big stable release built from 3,028 pull requests by 454 contributors. The headline is Advisor, an experimental, opt in feature that checks an agent's work while it runs and suggests a correction when it finds a clear wrong turn, like repeating an approach that already failed or drifting from what you asked. You turn it on in Settings, then System, then Labs. OpenClaw is upfront that it uses extra model calls, can miss a problem and can suggest a change you don't need. Advice does not run tools or approve actions, so you are still the human in the loop.

Also new: a Details panel that opens beside wide chats with session info and task progress, and a split speech mode for browser Talk that uses the same model as your typed chats, with separate services for transcription and speech. OpenClaw also says the new Particle-1 coding model is available once you update. If you self host, pin your version and read the notes before you upgrade, per our guide to using open source with your bots. (The last OpenClaw patch was in our Oct 30 roundup.)

Claude: faster long conversations

Anthropic put message threads into public beta on the Claude API yesterday. Instead of resending the whole conversation every turn, the API keeps it on the server and you send only the new messages, so Claude starts answering sooner. A thread expires 5 minutes after its last request, per the Claude Platform release notes. Anthropic says it already uses threads in Claude Code, where the median time to first token on requests of 500K to 1M tokens is about 35% lower, and the median request in long conversations went from about 800 KB to 1.8 KB.

Claude Code v2.1.317 also landed last night. Per the Claude Code changelog, it now asks before a web or shell tool call sends your account email address to the network, and warns you when a new cloud scheduled task picks up connectors you did not ask for. Small things, but they are exactly the checks you want on a bot that runs long running tasks.

OpenAI: a test for everyday tasks

OpenAI introduced LifeVal, a new benchmark for the everyday tasks people most want a personal assistant to handle, like household admin, budgets, groceries, recipes, workout plans and government services. OpenAI built the task list from two large time use studies plus its own survey of 3,995 people, and says it has used it over the past year to push its models on browser use, computer use and connectors. OpenAI shared example tasks, not an overall score, so treat it as a yardstick, not a ranking. The best test is still your own: give two bots the same prompt and score them on accuracy and steering, or run a full bot vs. bot bake off.

Quick Facts

- OpenClaw's new Advisor is experimental and opt in, and it suggests course corrections without running tools or approving actions.

- Message threads on the Claude API are in public beta, and Claude Code saw about 35% faster first tokens on very long requests.

- Claude Code v2.1.317 asks before a tool call sends your account email address.

- OpenAI's new benchmark measures everyday personal assistant tasks; there is no overall score yet.

- Want to see how these bots stack up? Check the Weekly Bot Scorecard or browse the directory.

See you tomorrow.