Alibaba laid bare the 14-round problem that's bleeding AI agents dry

On May 20, at the Alibaba Cloud Summit, CTO Dr. Feifei Li unveiled something the company had been quietly working on for months: a full Agent Infrastructure panorama. At the bottom of that stack, right at the computing power base layer, sat ANOLISA, system software with an ambitious claim: let every agent run on ANOLISA. The move fits into a broader platform play, as Alibaba is building both the brain and the immune system for AI agents, according to its hybrid strategy revealed earlier this year.

ANOLISA stands for Agentic Nexus Operating Layer & Interface System Architecture. It is not a new standalone operating system. It is an upgrade path: a three-layer set of kernel-level and runtime-level modifications that sits on top of existing Linux distributions, Alibaba Cloud Linux, Ubuntu, and others, and reorients the OS around agent workloads rather than human users.

The core insight, from product owner Zhou Xu at the Agent Native Infrastructure Breakout Session that afternoon, is simple and provocative. For decades, operating systems have been designed with a single assumption: the user is human. Agents do not use screens. They do not type commands. They need millisecond responses, structured interfaces, and the ability to run 24/7 without human intervention. And they are terrible at navigating traditional Linux.

The 14-round problem

Zhou Xu illustrated the gap with a concrete example: having an agent deploy a Python service on standard Linux. A skilled engineer needs about five minutes. The agent needed 14 rounds of conversation. The first 13 rounds were spent exploring directory structures, resolving permission issues, handling dependency conflicts, and troubleshooting network configuration, environmental reconnaissance that an engineer can skip by glancing at a familiar system. Only in round 14 did actual deployment begin. This echoes a broader issue in the agent ecosystem: even sophisticated coding agents spend most of their time on enviornmental reconnaissance rather than the actual task, a pattern highlighted in the analysis of how Devin measures human hours saved.

Token consumption analysis showed that about 80% of tokens went into environment exploration and trial and error; less than 20% was spent on the actual task. This is not an edge case. Internal data at Alibaba Cloud indicates that agents consume three to five times more invocation rounds than humans, and the vast majority of that overhead comes from understanding the environment.

The root cause is structural. Human users interact with an OS through a perception-understanding-decision-execution cycle, using mouse and keyboard, tolerant of second-level response times. Agents, by contrast, need CLI and structured interfaces, require millisecond responses, and can be hijacked by prompt injection, they lack the instinctive hesitation a human operator might have.

Three layers, four advantages

ANOLISA's architecture is straightforward. At the bottom, a distribution adaptation layer that works with Alibaba Cloud Linux, Ubuntu, and other Linux variants, no replacement of the existing OS required. The middle layer is the system optimization layer, where kernel-level tuning for agent workloads happens: improved scheduling, memory allocation policies, and interrupt handling for high-density deployment. The top layer is the runtime layer, encompassing agent observability, runtime enhancement, token compression plugins, and a security protection system.

Above that sits the encapsulation interaction layer, where agents can interact with the OS using intents rather than exact command syntax. This is where Cosh comes in, ANOLISA's default shell, which accepts natural language instructions alongside traditional CLI input. Cosh translates agent intent into system actions, removing the need for agents to remember parameter orders or file path conventions.

The measurable benefits are fourfold. First, token optimization: a claimed reduction of over 30% through three mechanisms. Less thinking (built-in OS skills act as an environment map, so the agent does not need to explore the file system from scratch), less loading (a Skill file system exposes only the minimum information relevant to the current job), and less transmitting (input and output are automatically compressed). Second, agent management, with full-link observability and a Skill ecosystem integrated with Alibaba Cloud's Skill portal. Third, runtime enhancement, including kernel-level performance tuning for Python and Node.js workloads. Fourth, built-in security, using a three-layer defense-in-depth architecture. This kind of structured agent infrastructure is also what Meta bet on when it acquired Manus for its agent layer, as noted in the analysis of that deal.

Security as the foundation of autonomy

The security architecture is particularly interesting. It reflects a new reality: in the agent era, security is not just about protection, it determines whether you can trust an agent to execute independently. If the security boundary of an agent is uncertain, no enterprise will let it run autonomously.

Traditional software is deterministic: known input, known output. Rules and whitelists work. Agents are probabilistic: unknown intent, unknown action. The same prompt, reworded, can produce completely different behavior. ANOLISA's three-layer approach blocks risks before execution (prompt scanning, code scanning, skill verification), monitors during execution (security observability, structured event logs, compliance audit, intent identification), and provides OS-level isolation as the last line of defense, even if the first two layers are breached, the operating system acts as a safety net.

Importantly, the security layer is designed to be agent-insensitive: no additional tokens are consumed during security monitoring, and there is no cost increase. The system claims to provide zero intrusion to existing setups.

Kernel-level tuning for agent density

The kernel optimizations are where ANOLISA makes its most tangible performance claims. In a future where each enterprise employee might be equipped with 10 or even 100 AI agents, hundreds of agent instances may need to run on the same server simultaneously. Traditional kernel scheduling and memory allocation are not designed for this density. This density problem is exactly the kind of infrastructure challenge that makes or breaks multi-agent collaboration, as demonstrated in the RecursiveMAS approach to scaling agent teams.

Alibaba Cloud's performance figures: concurrent memory load performance improved by more than 200%, memory allocation efficiency greatly increased, and interrupt processing performance optimized by close to 10%. The overall effect is a 30% reduction in agent execution time, a 20% improvement in Bench scores, and a 10% reduction in cold start duration.

A platform play in an agent world

ANOLISA is open source on GitHub, under the OpenAnolis community umbrella. The community is positioned as the open source platform for building the Skill Hub and agent ecosystem. It is already available on Alibaba Cloud ECS instances under the name Alibaba Cloud Linux 4 Agentic and will soon cover the Simple Application Server product line.

The move is strategically significant. As agent orchestration platforms multiply, managed agents, open source frameworks, harness engineering, the underlying infrastructure layer remains fragmented. ANOLISA is Alibaba Cloud's bet that the winning platform will be the one that controls the OS layer, not just the application layer. It echoes the shift from feature phones to iOS and Android: not that feature phones were not good enough, but that touchscreen interaction required a new system layer. In Alibaba Cloud's telling, the agent era requires its own operating system layer too.

The question is whether the industry will standardize around a single agent OS layer, or whether, as with Linux itself, multiple players will coexist, each tuned to their own cloud. ANOLISA's open source nature and compatibility with standard Linux distributions suggest Alibaba Cloud is aiming for broad adoption, not just a proprietary lock-in. But the model clearly benefits Alibaba Cloud's ecosystem: ANOLISA is already integrated with Alibaba Cloud Simple Application Server agent applications and will be tied to other cloud products.

For now, ANOLISA remains a bold proposition: that the future of AI agents depends not on smarter models, but on a layer of the stack most developers forget exists, the operating system. If the 14-round problem is genuinely widespread, Alibaba Cloud may have identified an inefficiency that no amount of prompt engineering can fix. The company's broader push to integrate Qwen into hardware devices, from robots to cars, shows how deeply it is betting on this agent-native infrastructure, as covered in the report on Qwen powering 150,000 devices.