Sabitlenmiş Tweet
An Van
2.6K posts
An Van
@anvanvan
Building Agent River. Stop waiting for your Claude Code agents. Let them wait for you.
#buildinpublic | AI dev tools Katılım Aralık 2022
122 Takip Edilen532 Takipçiler
@nilag_dev So far I like the better security and easy setup. No exposed ports (stdio subprocess), curated plugin allowlist, OS level sandboxing.
It’s new so there are still things missing, but I think they will close the gap soon.
English
Claude Code did the Apple strategy:
Don't be the first. Be second but with much better quality and UX.
Guess what product is the first to this ...
Thariq@trq212
We just released Claude Code channels, which allows you to control your Claude Code session through select MCPs, starting with Telegram and Discord. Use this to message Claude Code directly from your phone.
English
@Palanikannan_M Building the same and with more focus on protecting human attention and keep you in deep work
x.com/anvanvan/statu…
An Van@anvanvan
1/ The Agent River TUI is split into three zones. Like a river. ☁️ THE SKY: Ideas waiting. Sessions queued for later. 🫧 THE SURFACE: Agents bubbling up. They need you. 🌊 THE CURRENT: Agents running deep. Don't disturb them. The layout IS the mental model!
English
pov: you tried cmux and realized tmux was right there the whole time
all i needed was "is my agent done?" and fast switching. not a new app.
runs inside ghostty. same tmux shortcuts. freakishly fast
English
Adding opus 1m support meant we revised the context warning in humanlayer from “40%” to “100k tokens”
Matt Pocock@mattpocockuk
Doing some experiments today with Opus 4.6's 1M context window. Trying to push coding sessions deep into what I would consider the 'dumb zone' of SOTA models: >100K tokens. The drop-off in quality is really noticeable. Dumber decisions, worse code, worse instruction-following. Don't treat 1M context window any differently. It's still 100K of smart, and 900K of dumb.
English
@trq212 @dexhorthy That works only when main agent is using general-purpose agent with the skill.
If using named agents, they don’t have the skill tool and can’t invoke skills themselves.
Which leads to the problem dex is referring to: duplicate instructions in skill and agent files
English
@dexhorthy doesnt this just work?
Eg “use the playwright skill for X” or “launch a subagent to use the playwright skill for X” or even “use a play write subagent for X”
the parent should include instruction to load the skill in the handoff?
English
soooo subagent skills…feels like subagent instructions will overlap with skill and ppl won’t know what to put where. I can only assume that this will eventually replace custom subagents
This also explains why they recently made it so subagents can’t invoke skills.
I still think We need a solution for instruction modules orthogonal from context forking - let me use a skill in either parent or subagent dynamically based on a prompt @trq212 !!
Eg “use the playwright skill for X” or “launch a subagent to use the playwright skill for X” or even “use a play write subagent for X”
Today I STILL have to maintain two copies of the instructions to be able to run in parent or subagent!
Lydia Hallie ✨@lydiahallie
Btw you can add `context: fork` to run a skill in an isolated subagent. The main context only sees the final result, not the intermediate tool calls It gets a fresh context window with CLAUDE.md + your skill as the prompt. The `agent` field even lets you set the subagent type!
English
The CLAUDE.md file is your foundation, optimize it regularly. This skill makes it more convenient to get some basic stuff right.
dex@dexhorthy
we've been trying a bunch of stuff. this one kinda works.
English
I've made the same observations. So still /handoff around 80k tokens is a good rule of thumb 👍🏻
Matt Pocock@mattpocockuk
Doing some experiments today with Opus 4.6's 1M context window. Trying to push coding sessions deep into what I would consider the 'dumb zone' of SOTA models: >100K tokens. The drop-off in quality is really noticeable. Dumber decisions, worse code, worse instruction-following. Don't treat 1M context window any differently. It's still 100K of smart, and 900K of dumb.
English
@freekmurze Doing the same with the MacBook when I’m traveling. Only using mosh+tmux for direct access to Claude Code and my Agent River TUI for attention management.
Recently discovered getmoshi.app
It’s the best mobile terminal experience for Claude Code I’ve found
English
I'm walking through the city with my Mac close in my bag. I'm using the Amphetamine app to keep it awake even though it's closed.
I can use Claude Code on my Mac via my phone to quickly do stuff.
We're truly living in the future.
English
@ZagZino @milesdeutscher I‘m building Agent River to solve this and protecting your attention
x.com/anvanvan/statu…
An Van@anvanvan
1/ The Agent River TUI is split into three zones. Like a river. ☁️ THE SKY: Ideas waiting. Sessions queued for later. 🫧 THE SURFACE: Agents bubbling up. They need you. 🌊 THE CURRENT: Agents running deep. Don't disturb them. The layout IS the mental model!
English
@milesdeutscher Has anyone been able to come up with a way around it yet?
English
I feel like claude code is fucking my attention span
I end up mindlessly scrolling in between prompts
Takes a lot of focus to stay on task
English
@mstockton I‘m building Agent River to exactly solve that parallelization while protecting your precious attention.
The goal is to have a mindful workspace with agents where you can context switch with intention. In your terms, not when agents screaming for attention x.com/anvanvan/statu…
An Van@anvanvan
1/ The Agent River TUI is split into three zones. Like a river. ☁️ THE SKY: Ideas waiting. Sessions queued for later. 🫧 THE SURFACE: Agents bubbling up. They need you. 🌊 THE CURRENT: Agents running deep. Don't disturb them. The layout IS the mental model!
English
It certainly has forced me into a different way of working.
Ever since 4.6, Plans take a whole lot more time. Which is awesome since if you do it right, you can one-shot a whole lot.
My adjustment has been towards parallelization:
1. Build a backlog of self-contained changes
2. Curate a rough plan for all of those (usually voice memo to transcript works well as a baseline)
3. Parallelize as many plan modes as you can manage based on that backlog
Honestly 2 is the hardest part and where the real work is
Also helps to use some terminal emulator that helps you organize the windows well. Yes yes I know there is conductor out there, etc. but I want the actual terminal. I use ghostty with some customizations.
What am I missing? What else should I be thinking of?
Miles Deutscher@milesdeutscher
I feel like claude code is fucking my attention span I end up mindlessly scrolling in between prompts Takes a lot of focus to stay on task
English
Claude Dispatch/Cowork is insanely cool and I'm having a lot of fun using it to automate tasks.
The main bottleneck right now is that Claude for Chrome isn't good at all. It's like watching a drunk 4 year old try to use a web browser.
English
@anabology @felixrieseberg Did you updated both to the latest? Mine started to work when I did
English
@felixrieseberg have never gotten remote control to work properly, and same experience with dispatch now -- literally just can't create a session or pair phone to pc
English
By popular demand, Dispatch can now launch Claude Code sessions. Ask it to build, make, or improve something!
To use it, update your Claude desktop app and make sure you have Code enabled.
English
@desert_mouse Cool, good luck. Let us know how it goes.
How do you approach that? Ralph Looping?
English
@anvanvan Nice, I'm actually working on porting OpenClaw to Elixir, will see how it goes
English
Codebase size reduced by 84%.
Side project TypeScript codebase was getting too big (slop) so I asked Claude to switch to Elixir.
More robust, and it feels like it’s easier for Claude to handle.
English
> More generally, the principle of "garbage in, garbage out" applies here. There is no world where you input a document lacking clarity and detail and get a coding agent to reliably fill in that missing clarity and detail. Coding agents are not mind readers ...
gabby@GabriellaG439
New blog post: "A sufficiently detailed spec is code" I wrote this because I was tired of people claiming that the future of agentic coding is thoughtful specification work. As I show in the post, the reality devolves into slop pseudocode haskellforall.com/2026/03/a-suff…
English
English
We're shipping a new feature in Claude Cowork as a research preview that I'm excited about: Dispatch!
One persistent conversation with Claude that runs on your computer. Message it from your phone. Come back to finished work.
To try it out, download Claude Desktop, then pair your phone.
English
An Van retweetledi
damn this is so good and encapsulates everything I've been seeing/saying in the last few months
- a spec that is sufficiently detailed to generate code with a reliable degree of quality is roughly the same length and detail as the code itself
- so don't review those things, just review the code at that point, if you care enough about that level of abstraction
- unless you're vibing side projects or prototypes (yes, even zero-to-one software), you ABSOLUTELY SHOULD care about the code at that level of abstraction
- you need to find SOME way to get more leverage over coding agents though, because just reading all that code is a pain, esp when a lot of it is slop
- the default/dare-i-say-decel way is to go back to "i own the execution, and give little things to the agent, check it along the way"
- the accel-but-safe-way is to find something - NOT A SPEC (the word "spec" is broken anyway) - NOT 3 INVOCATIONS OF AskUserQuestion - that lets you resteer the model *before* it slops out N-thousand LOC
gabby@GabriellaG439
New blog post: "A sufficiently detailed spec is code" I wrote this because I was tired of people claiming that the future of agentic coding is thoughtful specification work. As I show in the post, the reality devolves into slop pseudocode haskellforall.com/2026/03/a-suff…
English