@zazencodes: Escape your coding agent's dumb zone... Past ~50% context the agent quietly gets dumber and more expensive, even with token caching Watch out and try to compact before you get there #coding #ai #claudecode #codingagents #contextengineering

Alex
Alex
Open In TikTok:
Region: CA
Tuesday 23 June 2026 16:17:59 GMT
35653
1097
60
100

Music

Download

Comments

tiktokdumbdumbsuck
maxrax :
I pay for the 1M context window, I’m gonna use the full 1M context window
2026-06-25 00:12:24
44
chance.ickada
chanceickada :
I cannot stress this enough but I never use compaction. It doesnt work. I have better success having the agent create a handoff document, starting a new convo, and feeding it in along with the transcript for the previous conversation.
2026-06-25 15:31:07
23
bretflix
Bort LeBort :
Bro you’re still compacting in mid-2026? I’ve been running one session per project for a month without compacting.
2026-06-24 04:10:36
7
luangomes7594
Luan :
compact makes it read everything and think about what to keep, that burns a lot of tokens for me, i just make it log everything to continue in a new session
2026-06-26 23:04:43
1
fiveclawfivemai
FiveClaw :
Idk Opus 4.8 cooks still at 900k context 😅 Sonnet 5? Not so much.
2026-07-12 09:59:55
2
samsamsapp
SamManCan :
What UI/IDE is that?
2026-06-29 22:33:16
1
amazingsmazing
smazing :
I prefer to run a master plan and start new sessions between phases
2026-06-25 06:12:47
5
sparecrypto
Divine Spark Co. :
bypass permissions is crazy.
2026-06-25 11:42:08
1
xab878bax
Red. :
Its close to 40% that the agents start to dumb down.
2026-06-23 22:06:48
2
legomyeggoplz
legomyeggoplz :
Stick around 200k and don’t compact
2026-06-29 02:11:43
1
secret56688
Udontknowme :
compacting process itself is also eating up token
2026-06-24 04:29:17
0
cbiscuit_._
Christian :
It needs to auto compact
2026-06-27 02:30:43
2
jayremoh
jay remoh :
why do we still have such a basic problem???
2026-06-27 05:51:59
1
thskjt
j18f77fw :
claiming am agent gets dumb because it has more information is crazy 🤣 and sayig to compact to solve it is even crazier
2026-06-26 00:10:09
0
fard0269
fard :
I see around 30-75 percent is the sweet spot. I try to fill that initial 30 percent with extremely precise information then use that window for it to instruct any subagents or long running tasks.
2026-06-24 11:35:51
3
tep8863
TEP88 :
Just use a context tool like ctxpipe - it’s open source. Or Mem0 / memco etc
2026-06-24 09:01:51
0
elipeters609
Eli Peters :
if you're compacting then you're not developing. clear clear clear and claude
2026-06-26 03:03:40
2
ayoyoyoy
Alex :
Just use OMP
2026-06-25 03:04:21
0
maxshulov
Max Shulov メ'𝟶 :
You’re using Matt pocock definitions
2026-06-26 11:22:53
0
cuatrosheep
Trey :
Stylist six party party three you can't do that
2026-07-13 02:46:55
0
hhec.protect
HHEC | CFD Protect :
Or be smarter and use “dynamic context pruning” plugin and let the agent it self compact and delete what context it dosnt need …
2026-07-14 15:17:31
0
greygoreeroy
greygoreeroy :
Automatic handoff documents when you clear context. Also, use an ai relay with another cheap subscription account/model
2026-06-29 04:23:32
0
anthonyk836
a.k :
this has been my issue now after watching i understand why my consistency in prompt generation for images starts dropping off even when I remind it to stick to the style I ask it.
2026-06-27 15:49:39
0
v8uccuxyxtvubomp
v8uccuxyxtvubomp :
i will try on opencode
2026-06-26 18:55:54
0
jeremycampbell752
Jeremy Campbell752 :
I more interested in the DBZ style it came up with!
2026-06-25 18:04:00
0
To see more videos from user @zazencodes, please go to the Tikwm homepage.

Other Videos


About