Newbie Spellweaver
- Joined
- Oct 23, 2026
- Messages
- 14
- Reaction score
- 21
Some AI Agents Notes
A lot of people still seem unfamiliar with AI agents, so here’s a quick overview.
An AI agent setup has two parts:
Both matter. Use the harness to open your project folder.
Here are some affordable combos I’ve tried that work well:
Always set the model to at least high effort mode. That’s what I’d call the minimum.
You can also use pricier models like GPT-6.1 Sol or Claude Opus 5.5, which are better at development and at planning project structure.
Each harness has its own settings. The good news: you can ask a newer AI model to help you migrate from one harness to another.
4 Key Points
1. README.md & AGENT.md
README.md is a guide to the folder. I recommend writing it as an index that leads step by step into the details. Otherwise the AI may read too much irrelevant information.
You can have README.md files in different folders. I usually keep one at the root and one per project, each giving a short index-style summary that points to the details.
AGENT.md teaches the AI how to work in this folder. For example: “don’t commit after every tiny change” or “don’t change test code while you’re still editing code.”
If you use Claude, name it CLAUDE.md instead.
How I split them:
README.md: for both humans and AI
AGENT.md: for AI only
2. MCP
Mostly used to let the AI control other software, like Blender or Ghidra.
Warning: having GPT-6.1 Sol or Claude Opus 5.5 use Ghidra may break their terms of service and could get your account banned.
3. Skills
Skills package scripts and workflows. Since they follow fixed scripts and steps, they rarely go wrong.
Example: with an MDL-editing skill, I just give the name of an outfit.
The skill extracts the MDL and opens it in Blender for me.
When I’m done editing, the skill repacks it and puts it in the right place.
It makes the whole process much easier.
4. Memory
I’m personally against using memory. Most harnesses can remember things about your working folder so you don’t have to keep reminding the AI.
But it’s hard to know what the AI actually remembered. In my experience, I’d ask it to do a one-off test I normally wouldn’t run, and it would remember that.
Later tasks would then run through weird extra steps, and it’s hard to make the AI forget.
Also, if you switch computers or someone else works on the repo, that memory is gone. It doesn’t transfer.
So I keep memory turned off almost all the time. If something really needs to be remembered, I put it in AGENT.md or in the README.md of the relevant folder.
That’s what I’ve learned over the past year. I hope it helps!
If you have your own experiences or anything to add, feel free to share.
and others
Sessions
When a session reaches about half its usage limit, get ready to run /compact.
Compact a session no more than 3 times.
If the AI starts going off track, stop it early.
If it starts holding onto wrong memories it can’t let go of, just start a new session.
Subagents
If you know a task will run for a long time, have the main agent send it to a subagent to handle in the background, then review the result when it’s done.
To add new requests, send a message. Don’t change the subagent’s README or records while it’s working.
Subagents’ work must not overlap.
I give each subagent its own job folder. Everything outside that folder is read-only. If a subagent needs a file, it copies it into its own folder.
Subagents must not use git, and must not create, delete or edit anything outside their folder. All commits are handled by the main agent.
For websites and documents you use often, have the main agent prepare one shared reference file for all subagents to read.
Commands
Every command should have an expected run time. When time’s up, decide whether to keep waiting or stop it.
Task size
For models below the level of GPT-6.1 Sol or Claude Opus 5.5, don’t give a subagent more than about 1 hour of work (including command run time), which is roughly 2–3 checkpoints. Otherwise it’s likely to go off track, and that can pull the main agent off track too.
A lot of people still seem unfamiliar with AI agents, so here’s a quick overview.
An AI agent setup has two parts:
Code:
AI model: the “brain”
Harness: the tool that runs the model and lets it work with your filesBoth matter. Use the harness to open your project folder.
Here are some affordable combos I’ve tried that work well:
Code:
GPT-6 Luna + Codex
claude-haiku-5-5 + Claude
opencode + deepseek-v4.1-flash-free / GLM-5.3-Flash Free
zcode + GLM5.3 / GLM5.3 flash
GPT-6 Luna + VS Code GitHub Copilot (expensive, but great for code completion)Always set the model to at least high effort mode. That’s what I’d call the minimum.
You can also use pricier models like GPT-6.1 Sol or Claude Opus 5.5, which are better at development and at planning project structure.
Each harness has its own settings. The good news: you can ask a newer AI model to help you migrate from one harness to another.
4 Key Points
1. README.md & AGENT.md
README.md is a guide to the folder. I recommend writing it as an index that leads step by step into the details. Otherwise the AI may read too much irrelevant information.
You can have README.md files in different folders. I usually keep one at the root and one per project, each giving a short index-style summary that points to the details.
AGENT.md teaches the AI how to work in this folder. For example: “don’t commit after every tiny change” or “don’t change test code while you’re still editing code.”
If you use Claude, name it CLAUDE.md instead.
How I split them:
README.md: for both humans and AI
AGENT.md: for AI only
2. MCP
Mostly used to let the AI control other software, like Blender or Ghidra.
Warning: having GPT-6.1 Sol or Claude Opus 5.5 use Ghidra may break their terms of service and could get your account banned.
3. Skills
Skills package scripts and workflows. Since they follow fixed scripts and steps, they rarely go wrong.
Example: with an MDL-editing skill, I just give the name of an outfit.
The skill extracts the MDL and opens it in Blender for me.
When I’m done editing, the skill repacks it and puts it in the right place.
It makes the whole process much easier.
4. Memory
I’m personally against using memory. Most harnesses can remember things about your working folder so you don’t have to keep reminding the AI.
But it’s hard to know what the AI actually remembered. In my experience, I’d ask it to do a one-off test I normally wouldn’t run, and it would remember that.
Later tasks would then run through weird extra steps, and it’s hard to make the AI forget.
Also, if you switch computers or someone else works on the repo, that memory is gone. It doesn’t transfer.
So I keep memory turned off almost all the time. If something really needs to be remembered, I put it in AGENT.md or in the README.md of the relevant folder.
That’s what I’ve learned over the past year. I hope it helps!
If you have your own experiences or anything to add, feel free to share.
and others
Sessions
When a session reaches about half its usage limit, get ready to run /compact.
Compact a session no more than 3 times.
If the AI starts going off track, stop it early.
If it starts holding onto wrong memories it can’t let go of, just start a new session.
Subagents
If you know a task will run for a long time, have the main agent send it to a subagent to handle in the background, then review the result when it’s done.
To add new requests, send a message. Don’t change the subagent’s README or records while it’s working.
Subagents’ work must not overlap.
I give each subagent its own job folder. Everything outside that folder is read-only. If a subagent needs a file, it copies it into its own folder.
Subagents must not use git, and must not create, delete or edit anything outside their folder. All commits are handled by the main agent.
For websites and documents you use often, have the main agent prepare one shared reference file for all subagents to read.
Commands
Every command should have an expected run time. When time’s up, decide whether to keep waiting or stop it.
Task size
For models below the level of GPT-6.1 Sol or Claude Opus 5.5, don’t give a subagent more than about 1 hour of work (including command run time), which is roughly 2–3 checkpoints. Otherwise it’s likely to go off track, and that can pull the main agent off track too.
Last edited:
