3 points:
1. Organize your repository/project: separate specs, plans, runbooks, code, config/settings, data, ... and architecture. Each with an index generated by a deterministic tool.
2. Define the standard architecture and don't let agents deviate from it without your approval. Also with deterministic control.
3. Use different models to validate the plan and then the code. If model A creates the plan/code, use model B to review it.
Just tell codex to do the copy/paste for you. Stay logged in to chatgpt site and it'll figure out the rest like any other site. Also the web side has its own usage limits ;)
One thing you can try is to build specialised subagents to help tease out more thinking when you're stuck with a model with limited reasoning budget as all the codex sub models seem to be (compared to API) --- but I have no firm, hard evidence of this so treat it as guesswork on my part.
Let me explain.
Creat a subagent that is told to just mull about a subject. This can be a simple file.
Have the agent expound on it with its own thoughts and additions.
Now rinse and repeat: give the file to another blank 'reasoning' subagent; then another; and so on.
This will eventually reach a point where they are unlikely to add more real value.
I believe they've locked down temperature and other sampling parameters so you cannot tweak these to get it to consider more unusual approaches.
The absolute best approach is a second model that is of a different family. Try Deepseek, GLM, Qwen, etc. and have it add its own thoughts. As a bonus, these models do not hide their reasoning. Put that in a file also.
start with a $10/m OpenCode Go plan, get yourself a gateway like GoModel and put all your credentials there, and then use a vendor neutral harness like OpenCode (because the big labs are not interested in supporting other model families than their own, they hold you down for the sake of their profits)
I made a shared memory that put memories related to the conversation and topic in each turn with hooks in claude code, that way my agents always have frees memories, no matter if im in claude, cursor or grok, it always remember my stuff
Someone once said, "give your agents good tools and context" and get out of the way. I would add, "stay in the loop friends"
I don't use Oai/Ant products or models, open weight absolutist that I am, but I would focus on your terminal workflows, custom tools, and repo context like AGENTS.md as a starting place. For tools, focus on moving deterministic workflows to scripts they can use, instead of having them do them differently each time based on markdown. You'll save yourself a lot of tokens and strife in the process.
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...
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.
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).
This is good advice and I wish I had read it before today. I use the state of the art model (Opus 5.5) and telling it with words how to interact with my issue tracker and my repo still doesn’t work. Mostly because of its overly restrictive auto-mode guardrails and brittle command whitelisting, but also a lot of “oops, I didn’t read the doc that you clearly tell me to read at the top of agents.md, I’ll make an ineffective note so this never happens again (it keeps happening)” The solution I’m using now is various scripts that encapsulate almost everything the agents usually need to do and just whitelist those.
I've been working on our custom PR agent at work. The agent phase is minimal, read-only/no bash (except for the .review dir). Most of the effort is in deterministic context gathering with scripts (before), and then posting the lines from a comments.jsonl file (after). I do have to work a something back into the agent phase, such that it can validate the comments.jsonl, because it seems to be generating bad payloads.
It's an iterative take-give, take away as much as you can so the bot can focus on the core task, then give some back to improve the agent output as you find repeated mistakes.
3 points: 1. Organize your repository/project: separate specs, plans, runbooks, code, config/settings, data, ... and architecture. Each with an index generated by a deterministic tool.
2. Define the standard architecture and don't let agents deviate from it without your approval. Also with deterministic control.
3. Use different models to validate the plan and then the code. If model A creates the plan/code, use model B to review it.
Just tell codex to do the copy/paste for you. Stay logged in to chatgpt site and it'll figure out the rest like any other site. Also the web side has its own usage limits ;)
One thing you can try is to build specialised subagents to help tease out more thinking when you're stuck with a model with limited reasoning budget as all the codex sub models seem to be (compared to API) --- but I have no firm, hard evidence of this so treat it as guesswork on my part.
Let me explain.
Creat a subagent that is told to just mull about a subject. This can be a simple file.
Have the agent expound on it with its own thoughts and additions.
Now rinse and repeat: give the file to another blank 'reasoning' subagent; then another; and so on.
This will eventually reach a point where they are unlikely to add more real value.
I believe they've locked down temperature and other sampling parameters so you cannot tweak these to get it to consider more unusual approaches.
The absolute best approach is a second model that is of a different family. Try Deepseek, GLM, Qwen, etc. and have it add its own thoughts. As a bonus, these models do not hide their reasoning. Put that in a file also.
Thanks for the response! I like the idea of pitting the various model families against eachother.
A bit token expensive, though. Especially when it isn't part of the subscription.
start with a $10/m OpenCode Go plan, get yourself a gateway like GoModel and put all your credentials there, and then use a vendor neutral harness like OpenCode (because the big labs are not interested in supporting other model families than their own, they hold you down for the sake of their profits)
I made a shared memory that put memories related to the conversation and topic in each turn with hooks in claude code, that way my agents always have frees memories, no matter if im in claude, cursor or grok, it always remember my stuff
Care to share your setup?
Someone once said, "give your agents good tools and context" and get out of the way. I would add, "stay in the loop friends"
I don't use Oai/Ant products or models, open weight absolutist that I am, but I would focus on your terminal workflows, custom tools, and repo context like AGENTS.md as a starting place. For tools, focus on moving deterministic workflows to scripts they can use, instead of having them do them differently each time based on markdown. You'll save yourself a lot of tokens and strife in the process.
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.
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.
Ah good point on the open weights. I'll certainly be poking into deepseek soon. Any preference with the Chinese models? Tips?
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).
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
This is good advice and I wish I had read it before today. I use the state of the art model (Opus 5.5) and telling it with words how to interact with my issue tracker and my repo still doesn’t work. Mostly because of its overly restrictive auto-mode guardrails and brittle command whitelisting, but also a lot of “oops, I didn’t read the doc that you clearly tell me to read at the top of agents.md, I’ll make an ineffective note so this never happens again (it keeps happening)” The solution I’m using now is various scripts that encapsulate almost everything the agents usually need to do and just whitelist those.
I've been working on our custom PR agent at work. The agent phase is minimal, read-only/no bash (except for the .review dir). Most of the effort is in deterministic context gathering with scripts (before), and then posting the lines from a comments.jsonl file (after). I do have to work a something back into the agent phase, such that it can validate the comments.jsonl, because it seems to be generating bad payloads.
It's an iterative take-give, take away as much as you can so the bot can focus on the core task, then give some back to improve the agent output as you find repeated mistakes.