Hi All, I'm trying to get Open-Hands to run for multiple hours doing useful work. (I'm using the agent-canvas). The task I have has very low ambiguity so in fact almost all the information required to complete the task is available to the agent. I can construct a general plan by asking the system to generate the plan and write it to a file. This plan has multiple phases. Now I want the system to construct a tree of tasks and subtasks for a particular phase. The best I can get from my prompting strategy is something like: "generate a task list for phase 4" (which generates something like 10-20 tasks) and then say "implement the task list". Execution of tasks to implement a phase might be $200-$300 for gpt-5.5. (I'll switch to a cheaper model pretty soon.)

Have you taken a look at these docs? https://docs.openhands.dev/sdk/guides/agent-delegation We’d be happy to improve them if anything is unclear

Do we have subagents working in agent canvas?

I had the same result yesterday - tried to do a simple sub-agent scenario and the agent tried to create an automation, not spawn subagents

Yup just found it! Worked 🙂

Anthony, we'd also love your feedback on Agent Canvas as you start using it! It's a newer project and we'd love to understand what works well, what's confusing, or what you'd love to see

Dear Joe, I'm learning a variety of different things simultaneously, so sometimes it's a bit hard to distinguish what is what. But I would say the following. Generally, the agent-canvas is much better than the CLI interface, which I was using before. The main "bug" I face now is that the scroll-back of an individual conversation is retained forever. Since I run large tasks, I find that chrome dies one way or another when the render process reaches 6 GB or so. I hand copy the task list to a file regularly. When it dies, I restart everything and then paste the task list into the conversation, which the agent figures out and adds to the task list tap.

Really helpful feedback! Just to understand, it sounds like you didn't want to work locally within a directory, but instead have OpenHands start a conversation from an existing repo. Would cloning the repo to your local machine, then opening the repo as a directory (using the "Open a workspace" feature) work for you?

it added the subdirectories as separate workspaces, which I didn't want.

sorry this was confusing! you can either pick "this directory" or "all subdirectories"

Oh right, duh. This works, thanks. I also added a plan in the planner, although it isn't clear how that helps me. I assume that the plan is added as context, and it automatically updates the plan as it accomplishes tasks.

👍 makes sense! CC @hundreds-twilight-32833

Thank you sharing @abundant-intern-35931! I think we can make some improvements here 🙂