News for hackers. Stuff that matters.

A comment on Ask HN: How to get the most intelligence out of your agents?. Parent comment

Re:Ask HN: How to get the most intelligence out of your agents? by thechairman12 on Tuesday October 20, 2026 @03:12AM (#50162412)
I've been working on building up my AGENTS.md file over time. It has certainly improved things. My core point is that even without these advantages, chatgpt.com provides deeper insights with their $200/mo ai...

Something is up.

[ Parent | on HN ]
Re:Ask HN: How to get the most intelligence out of your agents? by verdverm on Tuesday October 20, 2026 @03:15AM (#50162436)
it's better to keep AGENTS.md files (plural) minimal and spread/nested in your projects. Skills/reference documentation help to, but be wary of context bloat (from markdown/skills/tool defs)

> Something is up.

This is why open weights win. Deep transparency and the permissionless opportunity for anyone to investigate/benchmark is of primal importance to me. I'm not keen to have corporate-government (oligarchy) gatekeepers.

[ Parent | on HN ]
Re:Ask HN: How to get the most intelligence out of your agents? by thechairman12 on Tuesday October 20, 2026 @03:17AM (#50162454)
Ah good point on the open weights. I'll certainly be poking into deepseek soon. Any preference with the Chinese models? Tips?
[ Parent | on HN ]
Re:Ask HN: How to get the most intelligence out of your agents? by verdverm on Tuesday October 20, 2026 @03:22AM (#50162479)
OpenCode Go for $10/m is an excellent way to kick the tires on open weights

Getting an OEM (DGX) Spark is good if you want complete control or an excellent learning environment. You can do much more than just run llms for tokens (which will be generated much slower). They have become much more pricey tho...

Kimi K3 is good at the high end, DS 4.1 flash is good, most flash models are sufficient (I typically start a task here), I like qwen 27B at the small end (dense is better than MoE at this size).

[ Parent | on HN ]
Re:Ask HN: How to get the most intelligence out of your agents? by thechairman12 on Tuesday October 20, 2026 @03:29AM (#50162524)
Yeah I'll have to figure out opencode, never used it before.

I care the most about quality. It can be more expensive, and slower. My problems with the AIs are always intelligence related/effort related.

So I'm not sure if those smaller models you listed will be effective, given that I am having to turn to chatgpt.com's Pro to get quality answers.

Interested in Kimi & DS

[ Parent | on HN ]