Original post
Garry Tan#20
Garry Tan#20Peter Steinberger proposes automated orchestration loops to replace manual prompting for managing AI coding agents
Matthew Berman demonstrated loop orchestration triggered by GitHub PRs.
3324.3K37112.9K2M
Sentiment
Many users praised explanations of AI coding agent loops as simple and insightful while others dismissed the articles as obvious AI-generated slop lacking value.
Pos
78.6%
Neg
21.4%
39 comments with sentiment.
Cluster Engagement
Posts from X
Most Activity
Most Activity
VIEWS33.5KBOOKMARKS355LIKES255RETWEETS16REPLIES22
Matthew Berman@MatthewBerman
wtf is a loop?
Peter Steinberger 🦞@steipete
Here’s your monthly reminder that you shouldn’t be prompting coding agents anymore.
You should be designing loops that prompt your agents.
3hViews 33.5KLikes 255Bookmarks 355
Adi Oran@OranAITech
@mvanhorn https://youtu.be/mR-WAvEPRwE?t=1042&si=9ovo730xQTK3Y0cz
Run long running agents.
23hViews 13.5KLikes 30Bookmarks 94
Matthew Berman@MatthewBerman
imagine loops with Mythos...
Matthew Berman@MatthewBerman
wtf is a loop?
3hViews 4.5KLikes 39Bookmarks 26
Matthew Berman@MatthewBerman
@steipete I made a video explaining loops:
Matthew Berman@MatthewBerman
wtf is a loop?
3hViews 1.7KLikes 15Bookmarks 31
Matt Rice@bossriceshark
@mvanhorn @garrytan Great read, very helpful!
With where I'm at, I zero in on this:
15hViews 2.6KLikes 11Bookmarks 9
Matthew Berman@MatthewBerman
Reposting the banger quote from @tomas_hk “You won’t need to be a power user to spend a shit ton of money on AI. You will just need your AI agent to be a power user to spend a shit ton on AI."
Loops
Matthew Berman@MatthewBerman
wtf is a loop?
1hViews 2.4KLikes 17Bookmarks 4
Gen Z Mind@gen_z_mind
@mvanhorn Double Loop :)
20hViews 381Likes 2Bookmarks 5
Alcreon@alcreon
@mvanhorn Worth clarifying: loops don't replace prompts. It just determines what the prompt has to prove.
I find this quite helpful
21hViews 625Likes 4Bookmarks 3
Shubham Sharma | AI & Tech@editxshub
@mvanhorn If I’d wanted to read an AI’s opinion on the matter, I would have just prompted an LLM myself.
14hViews 2.4KLikes 19
Matt Van Horn@mvanhorn
@OranAITech Eyes 👀
23hViews 5.4KLikes 8Bookmarks 1
Filecoin@Filecoin
@mvanhorn The loop is only as good as its memory.
State that evaporates between runs doesn't compound. Verifiable, persistent storage is what turns a loop into something that learns over time.
9hViews 522Likes 10Bookmarks 1
Matt Van Horn@mvanhorn
@dsiroky Agreed. But Peter says he never writes plans!
23hViews 4.3KLikes 7
Zach Abrams@zcabrams
@mvanhorn Love your content but this is so obviously written by claude it’s painful
10hViews 873Likes 11
Morgan@morganlinton
@mvanhorn @garrytan I knew you were going to write this! 😅
15hViews 1.3KLikes 5Bookmarks 1
Reefy.ai@reefy_ai
@mvanhorn TLDR; "A loop is a small program that prompts the coding agent for you, reads what it produced, decides whether it is done, and if not, prompts it again."
Just a heads-up: run this on a subscription, not API tokens - unless you enjoy financial self-harm
18hViews 1.2KBookmarks 2
Agentspan@agentspan
@mvanhorn The observability, auditing, and guardrails piece is what will make this practice actually work. We created a project (http://github.com/agentspan-ai/agentspan) to address this in a way where you just define a few python methods and pass in the model.
12hViews 210Bookmarks 2
Trash Panda 🦝@trashpandaemoji
Scrubbed through this a bit and I think you're in the right direction, but I think what Peter and Boris were talking about the outer loop.
/goal or /loop aka "ralph loops" would be the inner loop. The inner loop takes a task or end state and just prods the model to keep going until it fully completes the task. This takes care of models just kinda being lazy or not fully finishing the task or just generally trying to be done with their agent turn.
The outer loop, or the "loop that prompts claude" is more akin to your automation setup. The outer loop in this example is an agent that watches the github issues, and spins up another agent/thread with its own goal to triage, validate, fix, and respond to the issue.
To implement features, the outer loop would play the role of the human previously. Instead of the human looking at github issues. The agent would handle it, then spin up a thread/subagent with its own goal to handle that one specific issue.
The outer loop is the start of the composition of these loops. That's what they mean. They're beyond automating features at this point. They're automating systems.
2hViews 49Likes 1Bookmarks 1
David Siroky@dsiroky
@mvanhorn You can pretty much do that by designing a plan in grok build or codex and then /goal it so it doesn’t stop before it’s finished. Needs a decent plan
23hViews 4.6KLikes 3
Sheing Ng@sashimikun_void
@mvanhorn @GeoffreyHuntley ^
23hViews 4KLikes 6
Krish Ray@KrishanuAR
@mvanhorn If I’d wanted to read an AI’s opinion on the matter, I would have just prompted an LLM myself.
16hViews 905Likes 3Bookmarks 1
Load more posts