Omar | Supergent
@agenticist
Followers
67
Following
7K
Media
368
Statuses
2K
Orchestrator of humans and agents.
New York, USA
Joined October 2025
Armin dug into Claude Code's plan mode and found it's mostly prompt injection + UI scaffolding. His take: why not just ask the model to plan in natural language? Counterpoint: structured modes matter because (1) they're instrumentation, not just UX: you can't evaluate what you
I don't use Claude's plan mode. I made some potentially uninformed statements today about plan mode, so I learned what it does by looking at the code. This might or might not be useful to you.
0
0
1
I don't use Claude's plan mode. I made some potentially uninformed statements today about plan mode, so I learned what it does by looking at the code. This might or might not be useful to you.
lucumr.pocoo.org
What I learned about plan mode as a non-plan mode user who plans.
24
28
387
Github announces charging for self hosted runners on the day I migrated to self hosted runners. destiny
0
0
0
@nummanthinks vite test parallelism + schema caching + squashed a render loop took me almost a full day of working with GPT 5.2 on extra-thinking, but we forked react-grab and got it down to 2.2s to make an update via the Live Edits, note they state 5.8s avg on their page
2
1
13
A discipline is engineering if (1) there is no bound to the rigor that can be applied to it, (2) it operates under real constraints of cost and feasibility, and (3) practitioners are accountable for navigating the space between unbounded rigor and finite resources.
0
0
1
interleaved thinking lets agents think between each step before: agent thinks once, then runs all actions now: agent thinks → acts → thinks → acts turns out it's just 3 lines of code
23
22
534
@simonsarris The tech will get better and will move way beyond LLM. A full toolchain, with simulator, verifiers, real-time knowledge, logic engines, all talking via language.
0
1
1
use open source to open the source
0
0
0
@irl_danB i knew the paradigm shifted when i used a jasper* translator for code (mid 2022) that explained what it did, sent it to a cracked senior engineer (azn, whose code it was) and he was convinced i was totally acquainted with the syntax/ code he was referring to (i was not) *an ai
0
1
2
@kyleturman build worlds where people can play and wonder not boxes where ideas get stuck
6
5
50
gladly, you must read
sadly, you must code
0
0
0
all we do is ultrathink
Advent of Claude Day 12 - Ultrathink You can control how hard claude will think before giving you a response. "think" → 4k thinking tokens "think hard" → 10k thinking tokens "ultrathink" → 31,999 thinking tokens Just say the magic word anywhere in your prompt.
0
0
1