@MrMarans the documentation appears to be in the Mintplex-Labs/anythingllm-docs repository, looks like you'd need another PR over there which references this one.

@timothycarambat Any progress on getting this reviewed? I would really like to have this feature.

I just updated the pull request to work with version 1.8.2 that got released yesterday.
Changes in the past months broke my original modification.

I also made users be able to /reset when the workspace is in agent mode, and removed the /exit command. Wasnt able to use custom slash commands tho.

I still dont really understand why agent mode isnt engaged all the time, is it to stop people from using tools in a chat instance?

I still dont really understand why agent mode isnt engaged all the time, is it to stop people from using tools in a chat instance?

Because the agent mode doesn't use rag. It can only load and search files, but not give partial data (as far as I know)

I still dont really understand why agent mode isnt engaged all the time, is it to stop people from using tools in a chat instance?

Because with lots of models for local it leads to a waste of a ton of tokens and smaller models do very very badly here. So you get by far worse results, which is why it is own action (and with this PR you can have always on agents too)

Could there be a way of using rules written into the system prompt that would engage agent mode? Ie if the user says abc-123 then switch to agent mode?

Could there be a way of using rules written into the system prompt that would engage agent mode? Ie if the user says abc-123 then switch to agent mode?

That is what @agent does. If you are asking about inferring agent calling based on the semantic meaning of the user query, that is what the LLM should do with tool calling. We have our own universal tool calling so we don't have to wait for the user to update the chat template, but the tradeoff is that the model may try to call a tool anyway even on a simple greeting

@timothycarambat: you mentioned on #4569 that there are good reasons this isn't merged yet. I understand the limitations if users would accidentally enable agent mode just anywhere. But it seems enabling it per workspace as @MrMarans suggested makes a lot of sense.
In my case for example I have a workspace dedicated only to legal research which I have an MCP-server for. So whenever I'm in that workspace I only care about agentic mode and nothing else.

Or did you mean there are technical/architectural reasons for not merging this? Maybe add it as an "experimental feature" like live document sync and mobile?
Glad to hear your thoughts.