@parthknowsai: AI agent without an orchestration layer. #ai #Tech #LearnOnTikTok #techtok #EduTok

Jimbo :

that doesn't offer flexibility, training models is not an option for the vast majority of use cases I would suggest

JakeCondemn :

Full circle? Back to using one LLM to do everything again 😳

Sam :

I guess if it's not meant to be production used. because the security goes from the harness to the trust me bruh of the LLM lol. it should work fine until it doesn't.

Damon Danieli :

This is a very specific case of classification and routing. (In the paper) they were using LangGraph and a 3B LLM as a classifier. They showed that they could train a 3B model to do both. This does not imply we can get Claude/GPT at a fraction of the compute.

Mateus :

Does the paper address multi workflow orchestration? Just curious

pvafree :

theres a recent paper that explores removing the orchestrator above the ai and uses a small ai to run the ai. see attached.ni need to find that pwper. Yes — the screenshot gives it away. It is from the follow-up paper by the same authors. The paper you're looking for is: Compiling Agentic Workflows into LLM Weights: Near-Frontier Quality at Two Orders of Magnitude Less Cost Authors: Simon Dennis, Rivaan Patil, Kevin Shabahang, Hao Guo Published: May 2026 arXiv: 2605.22502

Ben :

how much compute to fine tune the model? might be good for well-defined workflows, but this doesn't create a flexible tool. and wouldn't it be very opaque for auditing?

Alo :

but workflow changes a lot

Churro :

What’s the link to the paper? How’s the new gig?!

Janani Subramanian :

this sounds too good to be true

ultra instinct vegito :

Darcy :

What’s the model so I can use it.

Cr :

Very interesting, would like to try, any links on how to setup and train it ?

_nobleheart__ :

Jargon salad.

GKDev :

Benefits of orchestration is predictability. Yes, in some cases it will decrease quality and increase costs, but it’s more predictable than fine tuning

Cryo Kinesis :

Generally speaking having a simpler design for models works better. Modular design (for ai) is generally worse. But it comes with other drawbacks

QuantumTruth :

makes real sense. doing the same here.

111 :

then you woke up

Paralysisdp :

I thought the the same but came to realise it wouldn’t work on non local models

_habesha ነኝ :

is this just a paper?

Pablo :

Can someone explain how this makes sense? Orchestration has to be designed around a specific workflow. There are millions of possible workflows does that mean you need millions of LLMs with this approach?

deepinthedebug :

Matching a top tier model on a single facet is such a wishful flex. One more paper about “I trained a super specific niche task and my model outperforms generic models that didn’t train on this task.”

vicvill :

wow. game changer

goylivesmatter :

this is smoke and mirrors. they are running and agent on an 8b model and claiming it is magic. Uh. Ok. So. don't run agent on 1tb big models. Like... it is a non sequitur that proves nothing.

Kevin Hoff :

Link to paper?