
Community Lead @MindoAI / Hermesmaxxer
Joined February 2025
- Tweets3.2K
- Following513
- Followers525
- Likes7.2K
AGENT vs SUBagent when to use each one
Here is how I see it:
>Agent: one loop, sees every step
>Subagent: a fresh helper, own context and tools
>Never sees your chat, brings back only a summary
Same model, very different bill.
So when do I reach for subagents?
Whenever the work splits into pieces:
- research and search running in parallel
- keeping my main context clean
- a cheaper model for each small job (Hermes routes one per subagent now)
50 SUBAGENTS FOR SIMPLE QUERIES (bruh)
Google tested 180 agent setups and parallel finance tasks got 81% better with a lead plus helpers.
Anthropic got 90.2% better research with Opus 4 leading Sonnet 4 subagents vs Opus 4 alone.
And when do I stay with one agent?
When every step depends on the last one:
- planning and step by step builds (multi-agent setups did 39-70% worse there)
- routine and scheduled jobs, same steps every day
- anything where two helpers can disagree (in Cognition's Flappy Bird example one drew a Super Mario background)
Also it's not cheap. One agent burns ~4x the tokens of a chat, a lead with subagents ~15x.
Anthropic's early agents even spawned
[Subagents read. The main agent writes.]
1
8
210
Anyone who runs agents see what yours actually did
>You open the session log
>Wall of JSON
>You scroll
>You miss the one error that matters
I don't want to read Cua Driver logs.
I want to see the session.
So I drew one real session as a manga:
-Agent: Codex + Cua Driver
-Task: X article -> folder
-Calls:41
-Time 19 min 41 s
What you catch in 13 seconds:
what it clicked(in the background)
what it could touch my Chrome, only after --grant
where it got stopped -> unbound JS refused
REAL LOGS, NOT MOCKUPS
Session Replay every step from the log is a panel, every error is a panel too.
I want it inside Cua: run a session, get your replay (design not final)
@trycua Do you think it is a good way to read reports of your sessions?
3
102
You are not going to know which one is GPT 6.1 Terra
Put myself in a dirt for these ones
>Both generated from Opus 5.5 Max prompt
>Different harness
>You have 0% chances to spot
So have you cancelled your Anthropic sub already?
66
Replying to @MiaAI_lab
Look men from 1 prompt on MAX REASONING as HTML I get, and was needed to one more time ask for normal preview
119
Everyone built agents. Nobody built desks.
>It borrows your laptop
>It grabs your mouse
>It asks for your password
So Cua built one:
>Any agent gets a whole computer
>Laptop, Mac mini, cloud, sandbox
>You take over anytime
10 LINES TO START ONE.
Desks are solved rn @francedot
Opus 5.5 Earshotted better than X
Made a new sound for this button
Can you hear it?
Made it with Earshot from @LexnLin teapawt.net/bot/status/21122525077…
3
194
My weekly limit, one chat:
▓░░░░░░░░░ tetris animation
▓▓░░░░░░░░ 8 materials
▓▓▓▓░░░░░░ breakable ice
▓▓▓▓▓▓░░░░ sound
▓▓▓▓▓▓▓▓░░ 60 fps render
▓▓▓▓▓▓▓▓▓░ this post
Every message re-reads everything above it.
My Grok Bot limits ran out in two days, and I got tired of it
So I found a way out of this situation and compiled tips in a single article
6
163
Replying to @avrldotdev
Perfect, Codex finalized installation of OpenCode on my laptop just 5 sec ago and I see your notification
Will test 100$ Anthropic credits for API
1
1
143







