Good comparison of the 2 approaches of agent harness. Vita AI[0] also picked the outside model. We used the E2B sandbox solution, and fortunately it supports volume, so we just create volumes for personal skills and organization workspace, and mount them to the sandbox. This elegantly solves the "file system" problem you mentioned.
For openclaw to become helpful, you have to connect it to your personal email, access to your file etc. All of these requires user's manual setup right?. I do not get the point of "batteries included". Installing it is not the bottleneck right? The official docs has detail procedures for all deployment options.
Good try. I had the idea before, but later realized all the agent ecosystem has already highly coupled with bash scripts including skills. Though bash is not a good programming language, it's indeed flexible for CLI apps. I also realized that what agent and skills really need is just a glue language, not a real programming language, so bash is not so bad in the case.
Glad to to see open source models are catching up and treat vision as first-class citizen (a.k.a native multimodal agentic model). GLM and Qwen models takes different approach, by having a base model and a vision variant (glm-4.6 vs glm-4.6v).
I guess after Kimi K2.5, other vendors are going to the same route?
Can't wait to see how this model performs on computer automation use cases like
VITA AI Coworker.
Exciting to see Anthropic validate the "AI coworker" direction. We're building VITA AI (https://vita-ai.net) with similar philosophy but for enterprise QA testing.
One key architectural difference: Cowork runs sandboxed VMs on your local macOS machine, but we run sandboxes entirely in the cloud. This means:
- True isolation - agents never touch your local files or network, addressing the security concerns raised in this thread
- Actual autonomy - close your laptop, agent keeps working. Like delegating to a real coworker, not pairing with an assistant
- Scale - spin up 10 test agents without melting your CPU
The trade-off is latency and offline capability, but for testing workflows (our focus), asynchronous cloud execution is actually the desired model. You assign "test the checkout flow," go to lunch, come back to a full test report + artifacts.
Different use cases, different architectures. But the broader trend feels right - moving from conversational assistants to autonomous agents that operate independently.
Working on an AI coworker (QA engineer) that can autonomously tests web app like your human QA colleagues. It has its own virtual desktop and runs as background agent.
[0]: https://www.vita-ai.net (AI agent for content creation & social media management)