ANOLISA 1.0 Puts a Person and an AI Agent in the Same Linux Shell Session

Alibaba Cloud’s OpenAnolis community has announced ANOLISA v1.0, an operating-system layer meant to give AI agents the same formal standing on a Linux machine as a human user. The October 10, 2026 announcement describes ANOLISA as a “system housekeeper” for agents that sits on top of Alibaba Cloud Linux rather than replacing it. Version 1.0 is built on Alibaba Cloud Linux 4.0.3.0 and is active in the Alibaba Cloud Linux 4 Agentic Edition. The project’s components are developed in the open on GitHub under the agentic-os-org/ANOLISA repository, which carries an Apache 2.0 license and lists Linux and macOS as platforms.

The project name expands to Agentic Nexus Operating Layer & Interface System Architecture. Its README frames the goal around three practical constraints of running agents on servers: how they enter the terminal, how much they spend on tokens, and what execution environment they run in. Each capability can be turned on independently, so teams can keep the shell, agent framework and sandbox they already use.

One terminal instead of two

The headline feature in v1.0 is that people and agents now share a single command-line context. The OpenAnolis team describes a familiar problem: an agent running on the same machine as a developer typically works in its own session, so the developer cannot see what it changed or step in without losing the context it built up.

ANOLISA’s shell copilot, cosh, addresses that by keeping the user’s existing bash or zsh untouched and running every agent step in the same visible session, where the user can take over at any time. When a command runs normally, cosh stays out of the way. When a command fails, it steps in to explain whether the problem is a permission issue, a typo or a command that does not exist. The release also adds batch enable and disable for hook management and improved display of model authentication. Users already on the Agentic Edition move to the upgraded shell with a single cosh-switch command.

Skill security enforced below the agent

The second change moves security for agent Skills, the reusable task packages that agents load, from prompt-level reminders into the system. The announcement argues that rules an agent is merely told to follow cannot stop an agent that tries to get around them. In ANOLISA v1.0, the system decides which Skills an agent can see, which stay hidden and which are pinned to a trusted earlier version.

Installing or changing a Skill triggers an automatic rescan. When a threat is found, operators can block it outright, roll back to a trusted older version, approve it once after manual review, or trust it long term. The underlying components, skillfs, AgentSight and AgentSecCore, gained a smart activation mechanism that shows, hides and rolls back Skills according to security policy, support for Codex CLI, and a daemon for real-time monitoring. A bundled security dashboard places model calls, tool executions and security blocks on one timeline, so an operator can see the exact step where an action was stopped.

Measuring token savings before trusting them

ANOLISA’s Tokenless component compresses tool output before it reaches the model. Version 1.0 adds a comparison mode that runs the same session and task twice, once with compression off and once on, and reports token cost and savings per step and overall. The aim is to let teams verify that quality holds before enabling compression permanently. The announcement says real-world tests on mainstream coding sessions in Qwen Code cut wasted tokens by 30 percent, and Qwen Code support now works without manual configuration. That figure comes from the project’s own testing.

Snapshots and portable memory

The ws-ckpt component, which checkpoints a workspace, now supports scheduled snapshots at a chosen interval and a preview of file differences before rollback, plus plugin install and uninstall. The release includes one breaking change: the snapshot identifier parameter is now -s or --snapshot everywhere, replacing -i, which remains as a hidden alias.

The agent-memory component lets the context one agent has accumulated, such as project structure, user preferences and past pitfalls, be packaged and migrated to a different agent or model. When several agents run at once, their memories are isolated by default and can be fenced off or filtered. The component adds hybrid semantic search, automatic memory consolidation, privacy masking and cross-session task persistence.

Agents as System Tenants, Not Borrowed Shells

ANOLISA is part of a broader push by cloud and OS vendors to treat agents as first-class system tenants with their own permissions, audit trails and rollback paths, instead of scripts borrowing a user’s shell. The project’s README advertises alignment with OWASP’s Agentic Top 10 categories, listing seven as fully addressed and three partially.