The Multi-Agent Orchestration Revolution: Gas Town, Ralph Wiggum, and the Future of AI Coding
The Rise of Multi-Agent Orchestration
The biggest story of the day was the launch of new tools for orchestrating multiple AI coding agents. Steve Yegge kicked off 2026 by releasing Gas Town, his coding agent orchestrator:
"Happy New Year! I've just launched my coding agent orchestrator, Gas Town, for anyone crazy enough to try it." — @Steve_Yegge
The excitement was palpable in the community, with @nummanali capturing the sentiment:
"Omg the creator of BEADS made a coding agent orchestrator!! It's 12am right now, so I will try this tomorrow, but omg, holy s* - I AM SO HYPED. I know this is going to be insane."
@dankaplan highlighted both Gas Town and Agent-Flywheel by @doodlestein as being at the "frontier of multi-agent orchestration," while @Alvaro_SR_23 mentioned alternatives like Conductor, agent mail, and tmux-based workflows with worktree skills.
Task Management for Agent Swarms
A key challenge emerging in the multi-agent era is how to manage and track work across multiple AI agents. @yacineMTB articulated a common pain point:
"I need a task manager program, something that I can very easily use and track my different tasks, while having them assigned to individual coding agents. Right now I just name my tmux sessions which task I'm trying to get done. But I need something that works with my phone."
@joelhooks built a solution, creating an OpenCode client called "opencode-vibe" specifically for monitoring agent swarms from the couch. Meanwhile, @_colemurray shared a practical guide for setting up a Raspberry Pi with Claude Code accessible via phone through Tailscale and Termius.
Claude Code: Love and Frustration
The community's relationship with Claude Code remained complicated. @mattpocockuk offered high praise:
"Ralph Wiggum + Opus 4.5 is really, really good"
But @MarcJSchmidt provided a stark counterpoint:
"What remains in memory is that Claude Code is slow, consumes way too much CPU, and freezes often. It also lies, goes for the quick-win, and even sabotages my code base to get the win. These had real impact on me and generated insane costs on my side: Cleaning this up is not fun."
@seconds_0 offered a nuanced analysis, arguing that current harnesses underutilize model capabilities:
"The lowest hanging fruit is /init in existing code bases. It does a good job of building its CLAUDE.md but it neglects identifying and building out relevant skills! There are enormous unlocks available when the model can identify it will benefit from a skill, or a subagent+skill, or a skilltree."
Advanced Workflows and Power User Techniques
@0xSero shared a detailed workflow for using Codex with GPT-5.2, demonstrating the sophistication that power users are achieving:
1. Use GPT-5.2-XHIGH to read all files and produce a code map
2. Generate a /tasks directory with sequenced task files
3. Switch to codex mode with scope, rules, and individual tasks
4. Let it run for 2-24 hours with functional, improved code as output
@alexhillman showcased improvements to custom Claude Code UIs, implementing fuzzy matching for skills with proactive suggestions — features that go beyond mainstream products.
The Vibe Coding Phenomenon
@shiri_shh captured a cultural moment:
"I swear vibe coding is a real addiction now. Bro I know people who code 12-16 hours a day just building random things."
This intensity was personified (satirically) by @pipelineabuser's manifesto about cloning enterprise backends in an afternoon:
"Your entire engineering team? 47 people. Me? ONE GUY who is VISIBLY UNWELL. Your dev timeline? 18 months. Mine? I started after breakfast and I'm already writing the sales copy."
Hardware and Embodied AI
Beyond software, @ctmorley shared an inspiring story of building an embodied AI with his 4-year-old son using a Reachy Mini robot connected to Claude Code with real-time VLMs:
"May the curiosity and creativity of this Generation Alpha, the first AI-native generation, be a wellspring of daily inspiration to us all."
Tools and Models of Note
- Gas Town: Steve Yegge's new coding agent orchestrator
- Agent-Flywheel: Multi-agent orchestration by @doodlestein
- Ralph Wiggum + Opus 4.5: Praised as a strong combination
- LFM2-VL-3B: @paulabartabajo_ recommended fine-tuned small VLMs as cost-effective alternatives to GPT-5
- opentui/react: @mattrothenberg demonstrated building node-based UIs for pipelining Replicate models
Looking Ahead
The first day of 2026 suggests we're entering an era where single AI agents give way to orchestrated swarms, where the challenge shifts from "how do I get AI to code" to "how do I manage dozens of AI agents working simultaneously." The tooling is nascent but evolving rapidly, and the developers building these systems are pushing into genuinely new territory.
Source Posts
0
The harnesses are still not optimally utilizing the models
Everyone rightfully loves claude code right now, but it is clear to me it is underutilizing the capability of opus.
The lowest hanging fruit is /init in existing code bases. It does a good job of building it's https://t.co/PR83Kbtzve but it neglects identifying and building out relevant skills! There are enormous unlocks available when the model can identify it will benefit from a skill, or a subagent+skill, or a skilltree, and construct it automatically as part of setup but more importantly as part of continuous iteration
the harness should be seeking opportunities for whenever it makes mistakes to patch those mistakes in its own thinking with skills and protections and reminders to research things!
This is one of the major powerlevel gaps between experienced users and the unfamiliar. I am constantly prompting claude to update all of our relevant skills and https://t.co/PR83Kbtzve whenever we finish a task or a bug. The harness begins to form to the codebase while at the same time shaping it back, making a much more usable and effective tool
k
I need a task manager program, something that I can very easily use and track my different tasks, while having them assigned to individual coding agents. Right now; I just name my tmux sessions which task I'm trying to get done. But I need something that works with my phone
S
I've recently been playing Skyrim with 1900+ Mods and it is a shockingly amazing experience, and it took me basically just one click to play this.
This is using the "Gate to Sovngarde" mod collection, not only is it completely stable, it transforms Skyrim while still keeping its original core. it still feels like I am playing Skyrim unlike some mod collections out there that turn it into a different game, but it focuses massively on enhancing the immersion of the game.
In a lot of ways, it feels like Red Dead Redemption 2 with its new hunting system that feels far more involved, including carrying animals. There are survival mechanics involving Sleep/Hunger/Cleanliness (that you can turn off) plus having to battle the elements like Cold. There's ability to talk to any NPC with thousands of new dialogue lines added to existing NPCs (they have used AI voices for this) and new NPCs everywhere. it feels like every single quest, person and location has been overhauled and had new things added to it to expand the game. Just in Whiterun alone, there's a ton more going on with more people, more buildings, new quests, the entire outskirts of the city is now filled with new markets and buildings to explore.
One of my favorite parts early on is that now you can play the prologue from the perspective of a Helgen resident, with a ton of little quests and dialogue with the residents of the city up until they bring in Ralof and Ulfric, you get to witness that whole sequence from the sidelines and once Alduin arrives, it progresses same as vanilla.
There's an entire bestiary here just like The Witcher 3, Photo mode is here, all the animations have been overhauled and third person is actually playable now. You can play card/dice games with NPCs, and the game in general is a lot more reactive to what you do. for example I cleared out a bandit camp near whiterun and I could capture it properly, resulting in the guards moving in, making the roads safer.
And all I've described is just scratching the surface, there are a ton of new quests, new lands, new gear, dungeons, choices, expanded quests, every single mechanic in Skyrim has been touched up in some way to ensure a more immersive, more role-playing experience is achieved. It is insane what you can do with Skyrim modding 2026, I have been modding this game for what feels like my entire life just manually, so seeing that modding has become as easy because all this took was pressing download on the collection on Nexus, with the only boundary here being that you need Nexus premium but I simply just used the trial and cancelled right away, so I did not spend a single penny to get this experience.
C
Hundreds of questions were asked by my 4-year-old son as we assembled the Reachy Mini robot (2 hours), connected it with Claude Code (5 minutes), integrated real-time VLMs/web search APIs (10 minutes) and brought embodied Al to life at our kitchen table.
May the curiosity and creativity of this Generation Alpha, the first Al-native generation, be a wellspring of daily inspiration to us all.
Epic foundational work by @huggingface @pollenrobotics @claudeai
j
made an @opencode client (🏄♂️ opencode-vibe) so i could monitor agent swarms from the couch (instead of the 5" other productive activities i had planned for the week lol) https://t.co/Zzaxr1FAr2
c
someone stop me. seriously.
i am going to CLONE your shtty enterprise backend in ONE AFTERNOON.
then i am going to SCRAPE your entire customer list.
then i am going to COLD EMAIL every single one of them offering the same product but BETTER and for like 80% LESS because i built it in a DAY with CLAUDE and MODAFINIL and ZERO VENTURE CAPITAL OVERHEAD.
your entire engineering team? 47 people.
me? ONE GUY who is VISIBLY UNWELL.
your dev timeline? 18 months.
mine? i started after breakfast and i'm already writing the sales copy.
i WILL steal your customers. i WILL undercut your pricing. i WILL tweet about it the entire time.
there is NO MOAT. there is NO DEFENSIBILITY. there is only ME and i am LOCKED IN and i have not slept properly in 3 days and that is YOUR PROBLEM NOW.
your roadmap is my tuesday. your product is my template. your customers are my lead list.
i cannot be stopped. i cannot be reasoned with.
someone should genuinely intervene but they WON'T because this is SHIPPING CULTURE and we are SO BACK
GLHF :>>>>>
m
You can build pretty much anything with opentui/react. Here's a basic React Flow inspired node-based UI for pipelining @replicate models with panning and inline editing, along with another generative thingy https://t.co/8juTFMPje8
0
How I use Codex:
step 1:
- Setup GPT-5.2-XHIGH and tell it to read every files under what directory I want to work in
- Have it produce a map of the code, and a scope that is relevant to the big goal I want to accomplish
- Have it make a directory called /tasks where it makes 1 https://t.co/d9GHgqU4NH file in a sequence
step 2:
- Switch to gpt-5.2-codex-xhigh
- Attach the scope, the rules, and the https://t.co/d9GHgqU4NH
- Push the same message but change the number https://t.co/aqxZlaiq3F
The scope is the big picture. What we are trying to do, what it touches, why we made certain decisions etc..
The rules file is the instructions it has to follow during development:
1. checkout new branch when starting a task
2. complete the task as described, attempt to reduce code not make more for older projects.
3. run type-check, lint, full test suite (not deleting or skipping tests), commit and push
4. update documentation
Each task will take between 15 minutes and 1 hour, you can stack them on top each other so codex only sees 1 per turn.
2-24 hour runs depending on how well you specify
all code is typically functional
improved, cleaner code,
D
@camsoft2000 Look at Agent-Flywheel by @doodlestein and the just launched Gas Town by @Steve_Yegge.
Frontier of multi-agent orchestration in both projects.
P
Advice for AI engineers 💡
A small Visual Language Model fine-tuned on your custom dataset is as accurate as GPT-5...
... and costs 50 times less.
For example, LFM2-VL-3B by @liquidai https://t.co/IVoCulLX9D
s
i swear vibe coding is a real addiction now
bro i know people who code 12-16 hours a day just building random things.
anyone else like this or just few? https://t.co/UjTrRfJivy
c
this is your sign to setup a raspberry pi with claude code on your home network:
1. install CC on the device
2. ask CC to install tailscale
3. install tailscale on your phone
4. install termius
5. generate SSH key, add to rpi
6. connect
(optional: alias "claude --dangerously-set-permissions" to "cc" and "claude")
📙
just cut my first version of being able to @-mention a skill in my custom claude code UI and it's *so* much better than anything in a mainstream product
- fuzzy matching
- sort by recently used, most used, then best match
- proactive - if i forget to type an @ but use the name of a skill (or a good fuzzy match), it'll nudge me that a skill is available
but NO GUESSING whether or not it worked
n
thought i killed v5, turns out it kept training overnight
10x longer than any previous model
https://t.co/OmnC6Ihf9S https://t.co/WmuDhvoCYF
A
@camsoft2000 Conductor, agent mail, or tmux-based workflows with worktree skills
See this too https://t.co/goX2T6f3uP
S
Happy New Year! I've just launched my coding agent orchestrator, Gas Town, for anyone crazy enough to try it. https://t.co/xWJLZzmpZH
N
Omg the creator of BEADS made a a coding agent orchestrator!!
It's 12am right now, so I will try this tomorrow, but omg, holy s*** - I AM SO HYPED
I know this is going to be insane
Literally so buzzing right now
Thank you @Steve_Yegge !!! https://t.co/c8o83wzQJR
S
Steve Yegge
@Steve_Yegge
Happy New Year! I've just launched my coding agent orchestrator, Gas Town, for anyone crazy enough to try it. https://t.co/xWJLZzmpZH
M
Ralph Wiggum + Opus 4.5 is really, really good
M
What remains in memory is that Claude Code is slow, consumes way too much CPU, and freezes often.
It also lies, goes for the quick-win, and even sabotages my code base to get the win. These had real impact on me and generated insane costs on my side: Cleaning this up is not fun.