I'm experimenting with adding an info button on an agent response which exposes all Reasoning and agent metadata. I have an agent working now on pulling everything we can to expose it. Anything in particular people are interested in?

Thinking time and tokens?

Got those already, anything else?

This may be a little redundant, but which LM was used to generate the response? We can switch LMs now.

ollama-gpt-oss

Ah, sorry, I meant “show info about which LLM was used to generate the response in the info bar”

(Sorry the interject, but can't help it) -- I know one thing that varies is hte amount of information in the reasoning traces, with some models hiding it away or not exposing it. I think stuff we do to expose it where models provide it will be very helpful and a great advantage. Hoping this make sense

But we expose already everything or? I think it is a cool idea because we can make the chat more light. Do you have a video of how it works?

I'll make one later, we don't expose everything now, you can't fully audit each agent turn. So the idea is we expose everything—it's something other platforms don't do and is truly useful, especially figuring out when an agent got stuck or burned tokens on something unuseful. Consider this: we have a number of metrics in research already, I also made some new graphs from pulling data, inside of this dropdown we can show agent activity against other metrics in a database or json file. That way we can really see agent performance over time in comparison with other models.

(This is awesome, if you have the matrix as you do this, for what we added, versus what other popular coding agents don't have, that will be useful for me -- and the wider community)

I think that's where we can think about moat—we aren't incentived to hide conversation metadata from you.

Where this is a benefit most directly is obvious: reducing and predicting total agent cost per PR, Issue, Code line #, etc. When you can divide the total cost by the code length, you have a way to super-optimize and predict token spend across the whole pipeline. The predictive math is super valuable.

Not sure about that 🙂

@modern-fireman-99755 send me a screenshot.

if you are using agent-canvas locally you can find that in the folder

.openhands/conversations/<CONVERSATION_ID>

The agent-server is sending the agent answers to the FE and then it is also saving that locally, i.e., you could create very nice visual using that (for example tool usage, thinking, number of tool calls and so on) For cloud the history is a little different but same concept

@limited-student-10747 @modern-fireman-99755

Nice!

Can you add also cached token:)

Nice!