The name of the thread is provocative, but the premise is valid - I have yet to see anything produced by multi-agent frameworks (langchain or bespoke works) that produced value. Anthropic pushes vibeCAD, vibeVFX, vibePowerPoint but the results are underwhelming. The real value is in codegeneration and autonomous infra, research.
It really sounds like you’re doing it wrong (using multi-agent patterns of yesteryear).
The proper way is to use multiple agents for work involving very large context, and splitting the context amongst them. It effectively enables encapsulation and separation of concerns, which yields much clearer benefits when working at scale.
I have not tried deerflow 2.0 but the community showcase is not particularly impressive to be honest. Single agent system can do just as good. Any multi-Agent framework must constantly outperform single-agent on a different tasks.
Agree. We came to the same conclusion sometime last year. Now we solely focus on a single threaded agent - and make them run longer in a focussed way.
You dont need a team of humans to write great code - one engineer with solitude and focus is all it takes.
Snake oil is a bit strong, no? I would agree that the burden of proof is on multi-agent systems to show they are outperforming single-agent systems.
On my own evals I have seen this, though the improvement may not have been worth the extra cost.
The name of the thread is provocative, but the premise is valid - I have yet to see anything produced by multi-agent frameworks (langchain or bespoke works) that produced value. Anthropic pushes vibeCAD, vibeVFX, vibePowerPoint but the results are underwhelming. The real value is in codegeneration and autonomous infra, research.
It really sounds like you’re doing it wrong (using multi-agent patterns of yesteryear).
The proper way is to use multiple agents for work involving very large context, and splitting the context amongst them. It effectively enables encapsulation and separation of concerns, which yields much clearer benefits when working at scale.
I have not tried deerflow 2.0 but the community showcase is not particularly impressive to be honest. Single agent system can do just as good. Any multi-Agent framework must constantly outperform single-agent on a different tasks.