ANOLISA

ANOLISA is a system‑software stack that turns a Linux host into an “Agentic OS,” adding runtime support, security, observability and a token‑compression layer for AI agents. The provider highlights a reference test where token usage dropped by about 47 % after the tool's output is compressed before it reaches the model. Installation is performed through a single CLI command (`curl -fsSL https://get.agentic-os.sh | bash`), after which the components can be enabled on supported distributions without removing the existing environment.

The stack is divided into three conceptual layers. The User Entrypoint Layer offers the cosh‑ng terminal, which keeps your regular bash or zsh session while allowing you to type natural‑language requests. When you ask for a task, the terminal parses the intent, asks for approval if needed, and then runs the corresponding command, preserving the keyboard control for you. The Unified Stack Layer (AW) acts as a bridge between the agent framework you already use and the OS‑level services; adapters and plugins can be added so that the agent continues to operate with its original code base. The Agent Semantic Layer contains several utilities:

• AgentSight uses Linux eBPF tracing (requiring root or CAP_BPF + CAP_PERFMON) to record trajectories, LLM calls and token consumption, exposing the data for later inspection.
• Tokenless compresses tool output and simplifies schemas, so the model receives a smaller payload; the stack reports the estimated token reduction.
• SkillFS mounts a virtual filesystem that lists “Skills” – predefined OS instructions such as system administration or DevOps tasks, which agents can read on demand.
• Workspace Checkpoint (ws‑ckpt) creates a snapshot of the file system state; in a historical stress test the rollback completed in roughly 108 ms, allowing you to revert unintended changes.
• Agent Security Core classifies commands as risky and can enforce sandbox policies, preventing agents from executing potentially harmful operations.

Typical usage starts with installing the CLI, enabling the stack, and then invoking an agent through the cosh‑ng terminal or via SDK calls. Input can be a plain‑language instruction, a command line, or a request to read a Skill file; output may be the executed command's result, a compressed token payload, or a telemetry report from AgentSight.