OpenClaw 2026.10.5 makes your agents easier to manage

OpenClaw, the open source AI agent that people run on their own machines and servers, published version 2026.10.5 on the 18th of October 2026, according to the project's release page on GitHub. The update brings a new sidebar for agents, Stop buttons for helper agents, an Undo button for skills the agent learns on its own, and conversation links that anyone can open without an account.

The release covers 2,250 pull requests, meaning proposed code changes, from 637 contributors, the official release notes said. The OpenClaw account on X announced the same three headline features and the same figures in a post on the 19th of October.

A sidebar that keeps your agents in view

Many people run several agents at once, each with its own job. The new sidebar shows a longer list of agents as readable rows, the release notes said.

• Search: when more than six agents are set up, you can search by name or by ID, a short unique label.
• Pinning: with four or more agents, you can pin favourites so they stay at the top.
• Same names: two agents with the same name are told apart by their ID.

For an ordinary user, this means less scrolling and fewer wrong clicks when a team of agents grows.

A Stop button for the helpers

A subagent is a helper that your main agent hands part of a task to. Two new panels, Subagents and Processes, now sit beside the main conversation, according to the release notes. The first shows the work handed to helpers. The second shows recent output from commands the agent ran.

Each panel has its own Stop control. You can halt one helper or one command and stay in the same chat. The notes add that Stop now also reaches commands still running and follow up messages waiting in line.

For an ordinary user, a single runaway task can be cut short while the rest of the work carries on.

Undo for skills the agent learns

A skill is a saved recipe: a set of steps the agent can repeat the next time. OpenClaw can write these recipes by itself after watching a long piece of work. The tool that manages them is called Skill Workshop, and version 2.0 arrives with this release.

The release notes list what changes:

• You can see how a learned skill changed over time.
• An Undo button restores an earlier version.
• The library keeps up to ten saved versions of each skill.
• Unused skills are archived after 30 days and can be restored.
• Personal facts go to memory, and project knowledge goes to project notes, so skills stay focused on repeatable steps.

For an ordinary user, a skill that the agent learned badly can be rolled back in one click.

What background learning costs you

Learning runs in the background by default. The self learning guide says a review is queued once a chat reaches 10 rounds of model work since the last review. It starts after about 30 seconds of quiet.

Each review is one extra run on the same AI model and provider you already use, the guide said. That provider charges its normal prices for it. The guide gives these limits for each review:

• at most 2,048 tokens of written answer (a token is a small piece of a word),
• low reasoning effort,
• at most two minutes of running time.

The review reads a recent slice of the chat, capped at 32 KiB of text (about a dozen pages), and sends it to your model provider, the guide said. Most reviews make no change at all. A rate limit error, meaning the provider asks you to slow down, pauses all reviews for 15 minutes.

The release notes say a setting called Off stops background learning and the cleanup of unused skills. You can still ask the agent to learn something by hand in the chat. People who set learning to Off earlier keep that choice after the update.

Public chat links, and how to share one safely

A published conversation now opens at its normal chat address, and anyone with the link can read it without signing in, the release notes said. People who are signed in use the same link with their usual permissions.

The sharing guide lists what a reader sees: all your messages and the agent's replies, short summaries of the tools it used, and images. The agent's private reasoning and internal data stay hidden. Tool results stay private too.

Before you press Copy public link, the guide and the notes point to these checks:

1. Read the whole chat. Look for passwords, API keys (secret codes that unlock paid services), names and addresses. The guide says shared text is shown as written, with no automatic masking of secrets.
2. Remember that future messages are shared too. Everything you add to that chat later becomes public as well.
3. Start a fresh chat for sharing when the original holds private material, then publish that one.
4. Switch it off when you are done. Disable public access stops new readers. Copies people already saved stay with them, and preview images may stay cached for about five minutes.

Why this matters for an agent that never sleeps

Helper agents, background learning and shared links all depend on the agent running when you are away. A laptop that closes at night pauses every one of them. OneClickClaw hosts OpenClaw on a private server that stays on around the clock, with a Starter plan at EUR 14.99 a month and a 7-day free trial, according to oneclickclaw.io.