Trung Ng

Tech Lead, 6G AI ASIC, HW-SW co-design | Opinions are mine

[Trung Ng @trungnt13]

@HotAisle I recommended using Sol Medium, and setup the triage good enough so it escalate to high or xhigh.

It is faster than Grok 4.5, and code it generated very compact and straightforward. Occationally, admit cannot finish the task, but it left a perfect minimal scarfolding for xhigh

[Trung Ng @trungnt13]

New favorite workflow: Sol (med) for everything.

It will admit when it cannot do things and the "engineering" level is right.

Escalate to (xhigh) afterward, codebase is clear of over-engineering slop

[Trung Ng @trungnt13]

@chipro they do, they use batch jobs and subscriptions to increase utilization.

Everyone thinks sub is subsidized but how many of these subscribers have day work with enterprise accounts.

Compute market is dynamic since margin is low and providers wont absorb the dynamic

[Trung Ng @trungnt13]

@chris_j_paxton I never get this issue and work at very low level with custom SoC. Indeed, I would be happy if it "stop", most of the time, it goes over the line to get things done. I used both Cursor and Codex.
For sol, I deleted all old AGENTS.md and overwrite the harness instructions

[Trung Ng @trungnt13]

@morganlinton use tmux and just tell different harness to control each other, for example "ask cursor pane to review the pr" "ask cc pane to plan the changes" "ask codex pane to investigate the root cause" it is awesome and fast, no acp needed

[Trung Ng @trungnt13]

@theo less use doesn't mean useless, plan mode is used when it has highest leverage.
The number is getting closer to normal engineering workflow

[Trung Ng @trungnt13]

@MarcosHernanz "The agent is way more efficient at managing context, but it uses twice as many tokens."
no :)) what do you think where the "tokens" come from? the same token it generated will feed it context. Prime make muti-agents required, codex and cc forced this off since so many complaint

[Trung Ng @trungnt13]

:D

[Can Bölük @_can1357]

well this was a waste of tokens.
presented without commentary:

[Trung Ng @trungnt13]

@mitsuhiko It hallucinates a lot, I think part of it is from Google optimization to cut the reasoning and tooling short.

[Trung Ng @trungnt13]

@Yuchenj_UW they seems using pi harness very effective

[Trung Ng @trungnt13]

@yacineMTB only use SKILL.md and AGENTS.md

[Trung Ng @trungnt13]

@iScienceLuvr another fun use is tell it to gather and summarize all your chat from yesterday into an academic paper and print out for you daily

[Trung Ng @trungnt13]

@iScienceLuvr sign up new account with 20$ Plus, let Luna (max) with computer use exports every of your chat session to markdown

[Trung Ng @trungnt13]

Tried Terra xhigh, absolutely useless

Bask to Sol and Luna (max)

[Trung Ng @trungnt13]

Terra fast is slower than Sol?

[Trung Ng @trungnt13]

@pvncher tbh, it is much harder to debug multi agent session because OpenAI is encrypting all multi-agents commnuication

[Trung Ng @trungnt13]

@VictorTaelin its thinking trace is encrypted what you got is summary from other model

[Trung Ng @trungnt13]

@daniel_mac8 I have exactly the same setup and 1) no differences between Luna max, Terra max or Sol high for implementation, key is the exchange between Orchestrator and implementer; 2) Sol reviewer will always find over engineering issues that drive a pointless infinite loop. Simply not work

[Trung Ng @trungnt13]

@burkov take something open sourced, pi/codex/opencode, strip out the TUI and only keep the runtime, you will be happy

I prefer pi since it has the less built in features and no sandbox, permissions etc

[Trung Ng @trungnt13]

@simonw @dillon_mulroy I wonder if it use the same runtime as codex and some mcp, skills and agnetsmd added