# I've been building AI agents and I think we've been focusing too much on the brain — retained record

> Retained reddit record from reddit:artificial.

Canonical: https://highsignal.app/data/records/775ddcb0f3a02711

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:artificial)  published 2026-11-09 09:52 ingested 2026-11-10 02:37

[original source ↗](https://www.reddit.com/r/artificial/comments/1xrnopm/ive_been_building_ai_agents_and_i_think_weve_been/)

## retained text



```
so something ive been thinking about after building agents for a while is that for years pretty much all the energy in AI went into the brain, every model release was treated like a big deal and everyone was comparing benchmarks and context windows, and the assumption was kind of that once the model got smart enough the agent would just work on its own..

  
what i actually see now is that the models are close enough to each other that the model is rarely the thing going wrong, its the system around it, which is what people have started calling the harness. It's basically the system between the model and the real world that decides which messages reach the model, what information it gets at each step, which tools it can use, how those tools run and if the final reply is actually allowed to go out

a fun way I like to think about it is comparing it to the human body. the brain is the most expensive thing humans have and the whole body is kind of organised around not bothering it unless it has to, so you’ve got skin stopping most germs at the door, reflexes that get handled in your spine before your brain notices, red blood cells only sending oxygen where its needed, white blood cells checking for anything that doesnt belong and immune memory so you react faster the second time.

a harness usually starts with the same basics. a typical harness includes instructions, tools, context management, memory, permissions, verification and logging but what you build around those components is what makes some agents stand out from the rest.

also a fun side effect i noticed is the safer the agent gets teh cheaper it costs since so much of the token bill is the model reading the stuff it never needed or answering messages that should never have reached it in the first place.

curious how you guys are thinking about this, are you still mostly tuning prompts or have you started building more around the model ?
```



document fetched 2026-11-10 00:19 · canonical https://reddit.com/r/artificial/comments/1xrnopm/ive_been_building_ai_agents_and_i_think_weve_been
