Canvas implementation

[lspinheiro]

One challenge I have often using autogen for content generation with reflection using autogen is quickly exhausting the context due to the multiple content revisions that get added at each iteration. This is especially problematic for long documents which sometimes require small changes on each revision.

One possible solution to this problem would be to implement a Canvas where the agents can write/read directly to the canvas and only send messages with the comments while document changes are added to the canvas. Ideally, when creating the model context agents would be able to fetch only specific revisions of a diff history which would significally shorten context length.

[SongChiYoung]

I really loved that suggestion — super insightful.

One additional idea came to mind: since this approach is canvas-based, maybe we can actually use it to go beyond max token limits and let the LLM co-write documents that are longer than a single completion would allow.

I’ve been meaning to experiment with that kind of progressive long-form writing, so this could be a perfect starting point.

[lspinheiro]

I have a PoC #6215. I think it may work for long-form writing as well with the right systme prompts. Would love feedback. It is very simple right now and feedback to improve the design would be helpful.

[SongChiYoung]

@lspinheiro Absolutely happy to help! It’s a fascinating direction and something I’ve really been wanting to dive into myself.

[ekzhu]

I completely agree. Similar ideas have come before: workbench #4721 workspace etc. you can also think of MCP server as a stateful canvas with tools.

Glad it's finally being pushed forward.