relatedly, I'd love to get eyes on this: https://github.com/OpenHands/software-agent-sdk/pull/3213 It technically touches the prompts, but conditionally, so I don't think it'll change eval much

I like the idea of having controls, but suggested an alternative implementation method

asking openhands, but I'm not quite sure I understand the token usage problem current implementation (intended at least) is • description of all the controls goes into system prompt • current settings go into every user message (only a few tokens) ◦ for ACP we have to send with every message (see thread above this one) ◦ for non-ACP this would be better to only send when changed ▪︎ but IME the agent has a tendency to forget, so this might help

hmm, maybe the motivation for this PR wasn't very clear to me. why do we need this in the first place?

prior art

OK, but I don't like it, is there a way I can turn it off?

good question I was thinking that having default controls for everyone would actually solve a UX issue, where this behavior is undefined and changes from model to model; some LLMs like to commit, others will push; some like to run extensive tests once their work is done, others just make the changes

my worry is that this will cause the model to be less flexible in its ability to contextually make the right decision based on the task and cause unexpected behavior. to give an example in our slack integration we said "reply in the user's language" which seems very reasonable, but then this caused openhands to reply in Japanese when I asked questions in English but with a Japanese time zone or something

OK that makes sense!

I'd suggest that we have a space for

{personal_system_message}

or something in agent-sdk, and you can set this through

ConversationSettings

or

AgentSettings

(probably the former). Then if a UI wants to create sliders like the one you selected, the UI itself can auto-generate this message (or message prefix/suffix) and stick it in before starting the conversation. If you want adjust halfway through the conversation, then maybe it could be realized as a follow-up message appended to the conversation?

I want formal UI controls and settings for this--it's good metadata for us to have too

FWIW this sounds totally right to me

I'd suggest that we have a space for

{personal_system_message}

or something in agent-sdk, and you can set this through

ConversationSettings

or

AgentSettings

(probably the former). Then if a UI wants to create sliders like the one you selected, the UI itself can auto-generate this message (or message prefix/suffix) and stick it in before starting the conversation.

These controls seem a little too strict and fine-grained..., e.g. it's not obvious to me why I need one that does exactly the same thing as previous one, but also merges. I could just say "merge"? OTOH, it seems nice for a UI to have them and see if people like it, and send it to the LLM as a user message as suggested (I would love it if someday we could set

role=environment

in LLM APIs... alas today it's all "user")

Yeah, @wooden-umbrella-84098 I'm happy to discuss and brainstorm this more in a meeting, but TBH I think you're possibly overfitting a bit to your personal preferences and workflow. This seems like something where we'll need flexibility to fit various ways of working (and prompts/skills are a standard and effective way of doing that).

Fair feedback on overfitting. My goal is to give users more control via the UI--let user steer the agent via an explicit contract, rather than typing prompts like

do NOT push to GitHub unless I tell you to

or

KEEP POLLING GITHUB UNTIL THE TESTS PASS

I suppose the user could create skills for each type of agent behavior they want, and then trigger them with a slash command. But it's hard to e.g. switch from one strategy to another--you've now given the agent conflicting instructions

sorry, isn't that necessary anyway, since if we replace the instruction in message history, we break prompt caching?

the flow I suggest here is: • static system prompt with instructions for different settings ◦ e.g.

save=commit

means just commit locally ◦ e.g.

save=pr_ready

means keep iterating on the PR until it's mergable • initial user message (or maybe system prompt suffix) automatically includes

save=commit

or whatever the user's default is • if the user changes the setting in the UI, a follow-up message is sent saying

save=pr_ready now

With Graham's suggestion, I think maybe this flow can be similar? • regular system prompt • +`personal_system_message`, sent from client in agent settings ◦ maybe client sends exactly your instructions with definitions of modes ◦ and current default, e.g.

save=commit

◦ probably all added to system prompt suffix • if the user changes in the UI, UI generates a follow-up message saying ◦ the same, e.g.

save=pr_ready now

?

that seems like a good flow to me 🙂