I think it depends on the scale and quality of the data you have. Our project, General Evolution Architecture (Gear), starts with updating the harness and later the model weights and seed tasks for training (Under implementation). Please give it a try. We welcome any feedback.

Why do you use Jev to choose different items? I don’t believe Jev or any LLM today knows how to create an attractive landing page with great CTR and conversion. Jev can only speed up selection, but little performance upgrades to final outcome.

Maybe not a sandbox or VM I want, but a cloud computer. In this way, I don’t need to worry about storage, permissions and so on. It is for agents only.

A cloud computer is just a VM on someone else computer.

Why chose CLIP to do this. Have you tried small VLMs like Qwen-VL? I believe those models have video encoders can better perform at this scenario.

Not OP, but probably because they have no idea what they are doing or what CLIP even is. And the whole thing is likely just from one llm prompt, and then they dumped the whole thing onto GitHub, undoubtedly without even looking at any of the code or architecture.

And then you end up with image metadata processing code like this, which just by a cursory glance I’m sure has edge case bugs : https://github.com/allenv0/SCM/blob/main/screenshot-probe.js

Do you have task completion numbers alongside the reduction in tool calls? I’d be interested in whether narrowing the tool set ever removes something the agent needs later in a task. I’m building Gear, an open-source framework for evaluating and improving agent harnesses. One result that made aggregate metrics less reassuring: a harness change took our pass rate from 36% to 40%, but underneath that were 14 tasks improving and 10 regressing. I’d want to see that breakdown alongside the tool-call savings. The experiment is documented here: https://github.com/rsi-gear/gear/blob/main/docs/guide/en/exa...