ANOLISA is a layer Alibaba Cloud has built on top of Alibaba Cloud Linux — a "system housekeeper" for Agents, an operating system layer that gives Agents the same formal standing as a human user. It doesn't replace the Linux you already know; it adds native entry points, capabilities, and guardrails for Agents, so any Agent can run safely and efficiently on any machine. That's what we mean by an Agentic OS.
The recent ANOLISA v1.0 release is a milestone. This version puts people and Agents on the same CLI for the first time. It moves Skill security down from "remind the Agent to behave" to "the system enforces it." It takes token savings past "you can see them" to "run it both ways and compare exactly how much you save." It lifts snapshots from "you can store them" to "you can schedule them and preview before you roll back." And it closes a long-standing gap: you can now package up an entire memory and hand it to a different Agent to carry the work forward.
Let's start with the CLI, the core of it.
Scenario: You're working in the terminal and you also have an Agent running tasks on the same machine — installing dependencies, editing configuration, processing files in batches. But it runs inside its own session. You can't see what it changed or where it got to from your own terminal, and if you want to step in and take over, the context it built up isn't available to you.
Pain point: Even on the same machine, people and Agents used to work behind a wall. Either each side ran its own session with no shared state, or a chat box replaced the shell you know and you had to relearn every configuration, alias, and keyboard shortcut. People and Agents never had one shared terminal context.
Result: cosh in ANOLISA v1.0 puts you and your Agent in the same terminal context. The bash/zsh you know stays exactly as it is, every step the Agent takes happens in the same session in front of you, and you can take over whenever you want. Here, natural language and the command line are one and the same: when a command runs normally, cosh stays out of your way; only when a command fails does it step in immediately to tell you whether it's a permission problem, a typo, or a command that doesn't exist.
You used to pick between a CLI for people and a CLI for Agents. Now one CLI covers both.
Before: You installed a batch of Skills for your Agent — diagnostics, cleanup, inspection — very convenient. But you had no way to know whether one of them had been quietly modified, or whether an update had pulled in something risky. Worse, a lot of security measures amounted to "remind the Agent not to misbehave" — and if an Agent really wants to get around that, no prompt is going to stop it.
Now: ANOLISA moves that checkpoint down into the system layer. Instead of relying on the Agent to read the security rules and follow them, the system watches what the Agent is doing and blocks it when needed. The system decides which Skills the Agent can see, which stay hidden, and which are limited to a trusted earlier version. Install or change a Skill and the system rescans automatically — nothing to trigger by hand. When it finds a threat, you get more than "allowed" or "blocked": block it outright, roll back to a trusted older version, confirm manually and let it through this once, or trust it long term. Security isn't all or nothing here; you can see what's happening and choose. All of this maintenance runs in the background without interrupting the Agent's current work, which suits long-running workloads.
The security dashboard that ships with it puts the information out in the open. The overview page gives you conclusions, not a pile of data to analyze yourself. One click drills down from an abnormal signal to the specific event and on to the full chain. LLM calls, tool executions, and security blocks all sit on one timeline — so you don't just learn that a command was blocked, you learn which step it was blocked at.
Security can't rest on an Agent's good behavior. The system has to back it up.
The bill arrives at the end of the month and you want to cut token costs, but your first reaction is worry — what if the Agent gets worse once compression is on? And you have no idea how much you'd actually save. "Saving tokens" used to be a fuzzy account: you didn't know how much optimization actually saved, you couldn't say whether quality changed after you turned it on, so you were afraid to enable it and had no easy way to compare once you did.
ANOLISA v1.0 settles that account with a comparison mode. Before, you had to pick one: turn compression on without knowing the real savings, or leave it off and sleep easier. Comparison mode means you don't have to pick. The same session and the same task run twice, once with compression off and once with it on, and the system puts the two side by side: how many tokens each step cost, how much you saved on each step, and how much you saved overall. You don't have to gamble on quality — see the real savings first, then decide whether to turn it on, and switch back to compare any time. In Qwen Code, real-world tests on mainstream coding sessions cut wasted tokens by 30%. If you use Qwen Code, it's even easier: it works out of the box, with no manual configuration.
No guessing about token savings. Run it both ways and let the numbers talk.
The previous release's ANOLISA's snapshot feature got a warm reception, but users asked us this: an Agent edits a pile of files in one batch, you realize it went the wrong way and want to go back, and you never saved manually and can't remember exactly what changed. Now what?
ANOLISA v1.0 upgrades snapshots again, so your workspace saves and loads like a game. Set a schedule and it takes snapshots automatically at whatever interval you choose — no more worrying that you forgot to save, and every change to the workspace leaves a record. When you really do need to go back, you can preview the file differences you're about to recover, see what's what, and then act. No blind rollbacks.
With restore points you can return to at any time, experimenting costs far less.
You've probably had this moment: the Agent you're working with is finally dialed in — it knows your project structure, your preferences, the traps you've hit. Then you want to move to a stronger model, or hand the work to another Agent that specializes in it, and here's the problem: do you have to teach the new Agent all of that hard-won context from scratch?
ANOLISA v1.0 means you don't start over. The memory one Agent has built up can be packaged whole and migrated to another — switch models, switch to a more specialized Agent, the earlier context comes along, and the new Agent doesn't start from a blank page. And when you run several Agents at once, their memories stay separate by default and don't bleed into each other; you can also fence them off or apply filtered views as needed. Nothing leaks between them, and when you do need a handoff, the whole memory passes across cleanly.
Switch models without teaching from scratch, and run several Agents side by side without them stepping on each other.
Beyond the five user-facing capabilities above, ANOLISA v1.0 (based on Alibaba Cloud Linux 4.0.3.0) also ships matching upgrades to its underlying components:
| Capability | Component | Main changes in this release |
|---|---|---|
| One terminal | Copilot Shell（cosh） | Hook Management now supports batch enable/disable; improved model authentication display and keyboard interaction |
| Skill supervision + security dashboard | skillfs / AgentSight / AgentSecCore | The smart activation mechanism dynamically shows, hides, and rolls back Skills according to security policy; added Codex CLI support; introduced a daemon process for real-time monitoring |
| Token comparison mode | Tokenless | Added Qwen Code support; introduced dry-run comparison for the compression feature |
| Snapshot archive | ws-ckpt | Added plugin install/uninstall, rollback preview, and scheduled snapshots; incompatible change: the snapshot identifier parameter is now -s / --snapshot everywhere, replacing -i (-i is retained as a hidden alias) |
| Memory migration | agent-memory | Introduced hybrid semantic search and automatic memory consolidation; supports multi-agent memory isolation, privacy masking, and cross-session task persistence |
For the full list of changes, see the ANOLISA Release Notes:
https://www.alibabacloud.com/help/alinux/releasenotes
Enough talk — go type a couple of commands yourself. ANOLISA v1.0 is active in Alibaba Cloud Linux 4 Agentic Edition, and one cosh-switch command moves you to the upgraded cosh. Run a normal command first to feel how it stays out of your way, then deliberately type one wrong and watch it work out the problem for you. One try tells you what sharing a terminal with an Agent actually feels like.
Getting started:
https://www.alibabacloud.com/help/alinux/how-to-use-cosh-ng
SkillFS: Governing Dozens of Agent Skills with a File System
123 posts | 6 followers
FollowOpenAnolis - July 15, 2026
OpenAnolis - July 15, 2026
OpenAnolis - September 10, 2026
OpenAnolis - June 24, 2026
OpenAnolis - July 14, 2026
OpenAnolis - July 24, 2026
123 posts | 6 followers
Follow
Token Plan
Build more, spend less. One plan, every modality.
Learn More
QwenWork
QwenWork is dedicated to helping employees strengthen their professional competitiveness in the AI era and to enabling enterprises to improve organizational effectiveness.
Learn More
Qwen
Full-range, open-source, multimodal, and multi-functional
Learn More
Alibaba Cloud Model Studio
A one-stop generative AI platform to build intelligent applications that understand your business, based on Qwen model series such as Qwen-Max and other popular models
Learn MoreMore Posts by OpenAnolis
