In short: Alibaba has taken two of its relatively new open-source projects in the AI agent space and decided to develop them together, under one roof. This isn't just an organizational move – it's driven by the idea of building a unified infrastructure that covers both personal and enterprise use of AI agents.
Two Projects, One Family
Over the past month, Alibaba has open-sourced two projects with distinctive 'claw'-themed names:
- CoPaw – a personal intelligent assistant optimized for working with smaller models. The focus is on security, controllability, and stability. Simply put, it's an AI assistant you can trust with everyday tasks.
- HiClaw – an enterprise-level solution. This isn't just one agent, but a whole team: there's a manager who distributes tasks among worker agents, and this structure is designed for collaboration between humans and AI within organizations.
Both projects quickly gained popularity within the developer community. And now, the HiClaw repository has moved under the umbrella of AgentScope, Alibaba's framework for building multi-agent applications. Together with CoPaw, they are becoming part of a single ecosystem.
Why Join Forces?
When multiple AI agents work together, questions arise that simply don't exist with a single agent: How do they exchange context? Who passes what to whom? What happens if one of the agents gets stuck? How can you ensure an agent has the rights to perform a specific action, but no more?
The collaboration between HiClaw and CoPaw is aimed at addressing these very questions. The teams plan to work together on several fronts.
First is uniformity in agent construction. Currently, different agents may have various ways of describing their 'personality,' memory, and skills. The goal is to ensure that regardless of the 'intelligent engine' used internally, the external behavior of the agents is predictable and consistent.
Second is improving the collaborative experience. Long, multi-step tasks, work distribution among agents, state synchronization, context exchange – all of this requires robust mechanisms in a corporate environment. Add to that 'liveness' monitoring for agents and fault isolation, and it becomes clear that we are talking about non-trivial engineering challenges.
Third is system-level support at the infrastructure level. Corporate environments require strict control: who has the right to do what, how to track agent actions, and how to manage their lifecycle. Transparency is key here: interactions between humans and agents, as well as between the agents themselves, must be traceable and auditable. At the same time, humans retain the ability to intervene and adjust their work at any moment.
The Idea Behind It All
There is one principle that runs through this entire architecture and deserves special attention: the so-called zero trust model for agent access management.
Imagine an employee with a corporate badge – it confirms their identity but doesn't grant access to all doors at once. The proposed approach for agents is similar: each one has an 'ID,' but not a 'master key.' All calls to external services, language models, and tools pass through a single point of control. Security in this approach is ensured not by how 'well-behaved' the agent is, but by the fact that the infrastructure physically prevents it from going beyond its permitted scope.
This is an important conceptual shift: from 'trust the agent by default' to 'the agent operates within a controlled framework.'
The AgentScope Ecosystem: What's Included
AgentScope is not just a single tool, but a whole suite of components that cover different stages of working with agents:
- AgentScope – The core framework for building multi-agent applications.
- AgentScope-Runtime – Infrastructure for reliably running agents in a production environment.
- AgentScope-Studio – A visual environment for prototyping, debugging, and monitoring.
- AgentScope-Samples – A collection of ready-to-use examples and agent templates.
- Skills – A library of skills for agents in the ecosystem.
- Reme – A memory management framework for agents, supporting both file-based and vector storage systems.
CoPaw and HiClaw integrate into this ecosystem as specialized solutions: the former for personal use, the latter for enterprise use.
What About a Commercial Version?
In addition to the open-source code, HiClaw plans to leverage Alibaba Cloud's infrastructure, with the aim of turning these open-source developments into a ready-made enterprise solution. The commercial version of HiClaw is announced as 'coming soon,' with no specific dates given.
The project's authors specifically clarify that any products with similar names on the Alibaba Cloud platform are not the commercial version of HiClaw. This is an important disclaimer for anyone looking for the official product.
Overall, the developments around AgentScope represent an attempt to build not just a set of tools, but a cohesive infrastructure for serious work with multi-agent systems. Whether this approach will be widely adopted in practice, only time – and the reaction of the developers who use it – will tell.
