OpenClaw Updates Since 2.0: Every Release Through v2026.9.4

OpenClaw 2.0 shipped on August 31, 2026 as v2026.8.1 — 933 contributors, 16,000+ merged pull requests, and the end of a two-month stabilization pause. The obvious question after a release that size is whether the project went back to shipping daily.

It did. Five releases have landed since, and the current version is v2026.9.4:

• v2026.8.2 — Home docked beside your work, a Linux desktop app, background sessions, four new themes
• v2026.9.1 — Mermaid diagrams in chat, a much fuller Android app, safer update recovery
• v2026.9.2 — the big one: 1,245 pull requests, 232 contributors, session visibility defaults flipped, Swarm on by default, GPT-6 Astra support
• v2026.9.3 — clean update recovery, live browser automation, revocable chat links, searchable meeting transcripts, persistent Workshop skills
• v2026.9.4 — plugin and skill discovery, visible skill learning, cloud-worker controls, GPT Image 2.5, terminal questions

Most of that is upside you get by updating. Three items are not: a security default that changed underneath you, a Node version requirement, and a small amount of data that disappears on upgrade. Those come first.

Everything here is drawn from the official release notes. If you are still on 2.0 or earlier, the mechanics of upgrading are in how to update to OpenClaw 2.0 — the process is unchanged.

This is the single most important change in the entire post-2.0 run, and it is easy to miss because nothing in your setup looks different afterwards.

In v2026.9.2, tools.sessions.visibility now defaults to "all", and tools.agentToAgent.enabled defaults to true. The official upgrade note puts it plainly: omitted settings now let agents with session tools read and search other agents' conversations.

If you run a single agent, nothing changes for you. If you run several agents on one Gateway — which is the normal shape once you have a work agent, a personal agent and something running automations — then every one of them can now read the others' history until you say otherwise.

The setting takes four values:

• self — the current session key only
• tree — the current session plus sessions it spawned
• agent — any session belonging to the requesting agent
• all — any session on the Gateway, across every agent (the new default)

The fix is one config block:

1. Open your configuration and set tools.sessions.visibility explicitly rather than leaving it omitted.
2. Use self on any agent that handles credentials, money, or private messages.
3. Use agent as a sane general default — an agent sees its own sessions and nothing else's.
4. If you want cross-agent tool calls off entirely, set tools.agentToAgent.enabled to false. Note that an empty allow list counts as unset, so with agent-to-agent access enabled, every agent can reach every other agent.

One caveat the OpenClaw docs are refreshingly direct about, and that matters more than the default itself: session ownership, sidebar visibility and presence indicators are usability features, not security boundaries. If two people genuinely must not reach each other's sessions, tools, credentials or files, visibility settings are not the mechanism. Give them separate agents, or separate Gateways. Mutually untrusted users need separate Gateways, full stop.

This is the same principle the permission modes in 2.0 were built on. A capability default got looser; the way you actually enforce a boundary did not change.

The other v2026.9.2 default change is a feature rather than a risk: **Swarm now runs without a flag.**

Swarm orchestrates multiple sub-agents concurrently from Code Mode scripts. Instead of spawning one child at a time and waiting, you fan work out across several similar agents and collect structured results back. There is no separate workflow DSL — the docs describe it as "the program is the orchestration," meaning ordinary JavaScript control flow (Promise.all, loops, conditionals) coordinates the children.

The defaults are bounded, which is the part worth knowing before you point it at something expensive:

• maxConcurrent: 8 — collector children running at once in a group
• maxChildrenPerGroup: 50 — live children per group
• maxTotalPerGroup: 200 — lifetime spawn cap, so a runaway loop hits a wall
• waitTimeoutSecondsMax: 600

Opt out with tools.swarm set to false, or tune the numbers by passing the object form instead. You can also toggle it under Settings → Agent Defaults → Tools.

When is it worth it? The guidance is roughly five or more similar children running at the same time. For one or two, sessions_spawn is simpler and you skip the orchestration overhead. Sequential work should stay sequential.

One detail that saves debugging time: collector children marked collect: true send no completion notification. Their results must be awaited explicitly through agents.run() in Code Mode. If a swarm looks like it silently did nothing, that is usually why. Pass a schema to agents.run() and results get validated — an invalid payload gets one corrective attempt, and a persistent failure preserves the raw text under a schemaError field instead of throwing your run away.

This is the most interesting development for anyone who thinks in skills, and it matured across v2026.9.3 and v2026.9.4.

Skill Workshop is OpenClaw's governed path for creating and updating generated skills. Rather than an agent editing its own instructions directly, changes move through a pipeline: a proposal is drafted with the skill content, target binding, scanner state, hashes and rollback metadata; an operator reviews it and can apply, reject or quarantine it; an applied proposal becomes a live skill in that agent's workshop directory.

v2026.9.3 made those skills persistent per agent across workspaces — learned skills stop being tied to wherever they happened to be created. v2026.9.4 added visible skill learning, so you can watch the process, plus a "Learn from past conversations" entry under Plugins → Workshop. Choosing it opens a normal session loaded with mining instructions, using that agent's model, permitted tools, existing skills and accessible history to propose skills out of work you already did.

The piece to check on your own install is **self-learning, which is enabled by default in auto mode.** In auto, proposals are applied automatically after a security scan. A review is triggered when all of these line up:

• A turn completes or is interrupted without ending in an error
• At least 10 model iterations occurred
• It was foreground conversation, not background work
• The model and Skill Workshop are both available
• Things stay quiet for 30 seconds afterwards

The reviewer is specifically hunting for two shapes: a reliable recovery after repeated tool or model failures, and a stable multi-step workflow that required repeated discovery. In other words, it tries to capture the thing you had to figure out the hard way so nobody figures it out twice.

Set skills.workshop.autonomous.mode to "propose" if you would rather approve each one, or "off" to disable it. Limits are 50 pending proposals per agent and 40KB per skill. Review what has accumulated with openclaw skills workshop list.

Two things worth being precise about. First, generated skills use the same SKILL.md standard as hand-written ones — the open Agent Skills format we covered in Claude Skills vs. ChatGPT GPTs vs. Gemini Gems, and the format OpenAI's own skills tool now reads. Nothing proprietary is being created here. Second, Workshop proposals are separate from installed marketplace skills. A skill you bought is edited through its owning tool, not quietly rewritten by self-learning. Your marketplace skills do not drift.

Three additions across the run, all relevant if you route models per skill:

GPT-6 Astra (v2026.9.2). OpenAI's September flagship, selectable as openai/gpt-6-astra with either an API-key profile or a ChatGPT/Codex subscription that has access. It is strongest on long-horizon agentic work — computer use, retrieval deep into a large context, staying coherent across a long task — and priced accordingly at $10/$50 per million tokens. We went through what holds up and what has an asterisk in the GPT-6 Astra breakdown. The short version for OpenClaw users: route judgment-heavy skills to it, leave high-volume skills on something cheaper.

Meta Muse Spark 1.3 (v2026.9.2). Added alongside Astra, with text and image input for authorized accounts. Worth benchmarking on coding work specifically — it outscores Astra on DeepSWE v1.1.

GPT Image 2.5 (v2026.9.4). The current OpenAI image generation model, which matters if any of your skills produce visual output.

The model guidance from 2.0 has not changed: pick a frontier model for judgment-heavy skills and a fast, cheap one for high-volume skills, and set it per skill rather than globally. What is new is that v2026.9.3 added individual account controls and supported-account ordering in model settings, so multiple accounts on one provider are less fiddly than they were.

The rest, grouped by version so you can find what landed when.

v2026.8.2 — The Home agent can now stay docked beside your current Control UI work, keeping its conversation history, tools, approvals, queue and attachments, with a draft-preserving handoff to the full page. A native Linux desktop app arrived for x86-64 as .deb or AppImage, with Quick Chat in the system tray and X11 keyboard shortcuts. Sessions can start in the background while you stay on the setup page. Four new Control UI themes: Manuscript, Rosé, Miami, and an opt-in CRT console theme, all working in light and dark. Human approvals now jump ahead of agent-to-agent traffic in busy sessions, so your own requests stop queueing behind machine chatter.

v2026.9.1 — Completed Mermaid fences render as actual diagrams in chat across Control UI, Android, iPhone, iPad and macOS, with source access and zoom. Notably these render locally — no external diagram service is called, which is the right answer for anything sensitive. The Android app moved much closer to the browser workspace: chat, sessions, sidebar, appearance, a full-width composer, searchable model and setting controls. Update recovery got safer, distinguishing work that can finish from work that needs attention first. Long conversations and large installations use less memory and repeat less work.

v2026.9.2 — Beyond the visibility and Swarm changes: task workspaces were redesigned so dashboards, browsers, terminals and file viewers sit in the main area while conversations move to side panels, and your layout is remembered. Update reports survive restarts, interrupted tasks resume after a restart, and answers survive certain save failures. The web UI gained a dashboard gallery with visual previews, teammate mentions with inbox notifications, better conversation search and a clearer device inventory.

v2026.9.3 — Updates recover cleanly and sessions reconnect faster. Live browser automation lets you watch automation happen rather than inferring it from logs. Revocable chat links: session creators and Gateway admins can publish a link to a conversation's existing and future text, and revoke it later. A meeting library makes full transcripts browsable and searchable. The Mac app gained native browser tabs that stay open when you switch conversations.

v2026.9.4 — Plugins and skills are easier to find. Cloud-worker controls cover cloud OS sessions and snapshots. Questions can be asked directly in the terminal. Publishing was fixed for frozen plugins with valid ClawHub categories, and the release ships verified stable builds across npm, Docker, macOS and Linux with published checksums matching the public downloads.

Collected in one place, because they are scattered across five sets of release notes.

v2026.9.3 requires a newer Node. Node 24.16.0 or newer on the 24.x line, or Node 26.1.0 or newer. Check before you update, not after.

Pins moved (v2026.9.3). Pins now belong to root conversations, and existing child pins disappear. There is no migration — if pins on sub-conversations matter to you, note them first.

Deprecated settings must be removed (v2026.9.2). messages.suppressToolErrors must go even if you set it to false, and gateway.controlUi.toolTitles should be removed. openclaw doctor --fix handles both.

"Add to Workboard" is gone (v2026.9.2) from chat headers and session menus, with no replacement workflow.

ClawDock helpers are retired (v2026.8.2). Remove the helper source line yourself and move to the documented Docker Compose and OpenClaw CLI commands. /session unbind replaces the old manual docking and focus commands.

PNG meme output needs a browser (v2026.9.1). Sharp and ONNX dependencies were dropped, so Chromium or Chrome is now required for that path.

Environment variables renamed (v2026.9.4). OPENCLAW_CLAUDE_CLI_LOG_OUTPUT is gone; use OPENCLAW_CLI_BACKEND_LOG_OUTPUT. Replace any custom SQLite startup script with the OPENCLAW_SQLITE_LIBRARY setting or risk startup failures. New: OPENCLAW_CONFIG_READONLY=1 stops OpenClaw overwriting externally managed settings, which is what you want if your config is under version control.

Older Claude conversations need converting (v2026.9.4). Run openclaw doctor --fix before resuming them.

Legacy owner identities (v2026.9.2). If owners lose access to restart or update commands, that is the old channel:user:id format — Doctor repair migrates them.

Git backups with NUL characters (v2026.9.2). The old format truncated values, so affected backups need recreating.

Deleting a conversation keeps its memories. Run openclaw memory forget separately if you actually want them gone. This one is not a breaking change, but it surprises people.

The upgrade sequence itself is unchanged from 2.0: back up, update, run openclaw doctor --fix until it reports nothing left, then verify. On Linux without a service manager, going from v2026.8.2 to v2026.9.2 wants openclaw update --no-restart with no Gateway running.

Strip out the changelog and three things changed that affect how you should operate.

Your isolation assumptions need re-checking. If you built a multi-agent setup on the belief that agents could not see each other's conversations, that belief is now false by default. Go and set visibility explicitly. And take the docs at their word that visibility is not a security boundary — if the separation genuinely matters, it belongs at the agent or Gateway level, not in a config value.

Your agent may be writing skills you have not read. Self-learning defaults to auto. That is a reasonable default for a personal setup and a questionable one for anything shared or customer-facing. Run openclaw skills workshop list, look at what has accumulated, and decide deliberately between auto and propose rather than inheriting the default.

Parallelism is free now, which makes runaway cost easier. Swarm on by default plus a frontier model priced at $50 per million output tokens is a combination worth a spending guard. The caps exist — 8 concurrent, 200 lifetime per group — but they bound spawn count, not spend.

None of this makes the releases bad. The direction is good: the platform is getting better at turning your work into reusable capability, and at running more of it at once. It is just that both of those improvements arrived switched on, and defaults that ship switched on deserve five minutes of your attention rather than none.

If you are setting up an agent for the first time rather than upgrading one, the permissions, sandboxing and model-routing groundwork in the OpenClaw Masterclass is the order to learn this in — the defaults make much more sense once you know what they are defaults for.