Stijn

Finally got GPT-5.5 on Codex to think way harder than it normally does. As one would expect it to be 'think very hard', 'ultrathink', 'deep dive' etc - does not work. A simple 'use at least 100 (onehundred) tool calls before coming up with an answer' works really well

@badlogicgames @mauricekleine @SamuelBeek @lucasmeijer @ssh_exe_dev fantastic, thanks!

@StijnSmits @mauricekleine @SamuelBeek @lucasmeijer there is only @ssh_exe_dev

new electronics project. don't uave the time to design my own littlr AI robot. still waiting for my HF robot. so i'm going to frankenstein this €9 one and give it a brain. wish me reversing luck!

@badlogicgames @mauricekleine @SamuelBeek @lucasmeijer And also, we can try to add custom support right now for your gear, as we e.g. did for the (Claude Code) cardputer this week. Feel free to give any recommendations ('consumer-hardware-focussed') of something we don't support yet (officialy) but can make readily available for you

@badlogicgames @mauricekleine @SamuelBeek @lucasmeijer curious, at what stage would you add custom code execution for Schematik? We can do a LOT of stuff through sophisticated, extensive tooling imo, but maybe this is still not exhaustive enough. What platform would you recommend, daytona/Vercel or self-hosting such a thing?

@badlogicgames @SamuelBeek @lucasmeijer :D

@SamuelBeek @StijnSmits @lucasmeijer i'm very very glad :D

@badlogicgames @SamuelBeek @lucasmeijer Yes! our incoming agent-runtime.ts creates a Pi Agent, registers Schematik tools, and uses Pi lifecycle hooks such as transformContext, beforeToolCall, afterToolCall, and subscribe

@StijnSmits @SamuelBeek @lucasmeijer wait, schematik uses pi?! i'm equal parts happy and scares that it will spot out high voltage/amp schematics now :D

@badlogicgames @SamuelBeek @lucasmeijer Awesome Mario!! New agent version is coming soon (built on pi (!))

@SamuelBeek @lucasmeijer oh i hope i didn't come off to negative! digging into it as we speak!

@SaidAitmbarek thanks! also using it for 60 or 80 tool calls etc depending on complexity :)

@StijnSmits interesting heuristic approach ;)

Finally got GPT-5.5 on Codex to think way harder than it normally does. As one would expect it to be 'think very hard', 'ultrathink', 'deep dive' etc - does not work. A simple 'use at least 100 (onehundred) tool calls before coming up with an answer' works really well

@taka_start_up Curious if you have tried it!

@StijnSmits この発見、かなり興味深い 僕はClaude Codeで毎日開発してるけど、思考の深さ指定で出力品質が全然変わるのは体感してて。GPT-5.5でも同じ傾向出るなら面白い プロンプトの指示粒度がますます重要になってきてる🚀

@ghulamio Yeah it seems to reach at least 70-80 tool calls every time for me

@StijnSmits surprised system prompt can be gamed like that

@GolerGkA just a prompt, this is in between goal and normal thinking - if you have a specific problem you want to solve and get a lot of references for it, where goal would be overkill

@StijnSmits /goal or just a prompt?

@dea_tea_1 hai!

本当？あとで試す

@lfji Really interesting! Thanks

I dug into it, and it looks like it was actually conflicting with `/goal`'s internal instructions in the old 0.130.0-era implementation, not necessarily my AGENTS.md: #L15" target="_blank" rel="nofollow noopener">github.com/openai/codex/b… The latest version seems to have improved those `/goal` instructions quite a bit, and that wording is gone: #L19-L28" target="_blank" rel="nofollow noopener">github.com/openai/codex/b… So yeah, I should probably just update Codex lol. I had assumed it was some system prompt / AGENTS.md interaction, but I would not have thought to check this without your question. Thanks

@paulbettner that's a great one, thanks!

@StijnSmits another trick I've been using that works really well: 'before finalizing your answer, spawn parallel adversarial agents to attack your proposed conclusions from multiple angles'

@lfji This happens for some, can you ask if it had something in its instructions that blocked it?

@StijnSmits Tried it, but my Codex immediately went: "I won't make 100 tool calls just to hit a number, since that would be intentional work padding." lol

@dedene Because it’s an mcp

@StijnSmits Why not use the sequential thinking mcp?

@ileppane for sure! testing more tomorrow

@StijnSmits This has got to work!

@tawer1O yes

@StijnSmits Does forcing many tool calls raise real usage costs?

@slopwareindy this is different than a goal, try it out :)

@StijnSmits I do a similar thing with pi-goals, except it's built-in: x.com/slopwareindy/s…

@narrativenavi hahaha got to look into it, thanks for the reference

@StijnSmits I was jokingly applying the idea of Japan’s “karoshi line” to AI, as if AI were a human worker. Looks like that nuance got lost in translation.🤣

@narrativenavi arigato, what’s the Karoshi line?

@StijnSmits 過労死ラインを考えてほしい。 果たしてそれは最適な計算量なのか疑問があるぞ！ｗｗｗ