Sebastian Hirt

Searching

[Sebastian Hirt @Mokronoss]

@RhysSullivan @zeeg I am using executor sdk in my project and i think i had a similar problem where its annoying that the agent just doesnt know whats possible/available. I solved it by putting the source/integration name dynamically in the tool description of execute. Feels cleaner than extra tool.

[Sebastian Hirt @Mokronoss]

@RhysSullivan A similar thing is true for workflows. You can express via config (e.g. n8n) and have nice viz/no-code for non-tech users or use code (e.g. temporal or 'effect/workflow'), be more expressive, but difficult for non-tech users to review. Easy to write now though with agents.

[Sebastian Hirt @Mokronoss]

@maria_rcks I feel like we should "just" make agent runs durable. It should be doable to make the harness save all the streamed events and then send them without a new user message. Most harnesses save these already anyway to display the session again.
But benches would be nice anyway.

[Sebastian Hirt @Mokronoss]

@jxnlco Can you go into more detail about how much you guided the model on the process and/or what decisions the model/the team made? Like did you just give it a target loss and param size and gave it access to the training harness, datasets, rl and gpus? Or was it more back and forth?

[Sebastian Hirt @Mokronoss]

@meijer_s Ctrl-p --> disable paste summary

[Sebastian Hirt @Mokronoss]

@badlogicgames Isn't the provider round trip not an issue because of parallel tool calls?
The multi-edit is still probably more efficient and some providers might not have parallel tool calls.

[Sebastian Hirt @Mokronoss]

@ByronBernstein https://t.co/l4sK2CllMz :)

[Sebastian Hirt @Mokronoss]

@ForsenSC2 just order from like 5 different stores and cancel once one is getting sent .. worked for me .. got mine on the 17nd

[Sebastian Hirt @Mokronoss]

@ArchonAmazHS @ArchonHS 1st - Value Town, 2nd - Cloud 9, 3rd - Nihilum, 4th - Tempo Storm