3 ms·

This is pretty awesome! Just finished implementing into my own project that heavily relies on graphs, and am loving the initial results! :D

by gemisis 2mo ago

furkankly 2mo ago
Thank you, that's awesome to hear! Curious what the project is, if you can share. And if anything in the API felt rough, an issue would be appreciated.
- gemisis 2mo ago
  API seems to be fine! I'm admittedly using AI for the integration, so minimal exposure on my end, but agents seem to be good at navigating the API which is good. I'm using it for my project Leviath (leviath.dev), focused on building structured agents. In particular, the way that agents are constructed is via a graph, and so I have a TUI with this that allows both viewing runs, and creating agents with it now in a terminal nicely :D Previously only the web interface really was viable, now it's all doable via CLI which is nice for folks who just want to go through there.
  - furkankly 2mo ago
    That seems like a nice fit for rataflow and I'm really happy that it worked well for you! I didn't get to try it out, but Leviath seems like a really cool take on agent orchestration and context management. One question, afaiu, a single agent is composed of multiple stages where each one can possibly make an llm call to a different provider. Does that out-of-the-box multi-stage agent experience lead to the shared append log kind of context window given an agent? I also wonder how this compares to dynamic workflows from Claude Code. Thats one solution to the context bloat that I really love using.
    - gemisis 2mo ago
      > Does that out-of-the-box multi-stage agent experience lead to the shared append log kind of context window given an agent? Leviath supports doing transformations on the context window between stages (https://leviath.dev/docs/stable/stages#carrying-context-across-an-edge https://leviath.dev/docs/stable/stages#carrying-context-acro... has documentation on built in transformations, and scripting custom ones). so you could do a direct transform and then just append onto the region, or you could do something more complex like turn a list of items into a map of key-value pairs on some portion of the list. > I also wonder how this compares to dynamic workflows from Claude Code. I played around with these a bit, and what I found is that I just couldn't get the level of control I wanted. In most instances, I want to design a graph for the agent, and then have that graph be the execution for it (which is exactly where Rataflow comes in great :D https://x.com/GEMISIS/status/2090212911595872556 https://x.com/GEMISIS/status/2090212911595872556) The other thing is that dynamic workflows just seem very.... bloated. Sometimes I want a fairly simple flow, and workflows are just very heavy (whereas for heavier workflows, I haven't found them to give good results)

This is pretty awesome! Just finished implementing into my own project that heavily relies on graphs, and am loving the initial results! :D

by gemisis 2mo ago

This is pretty awesome! Just finished implementing into my own project that heavily relies on graphs, and am loving the initial results! :D

furkankly 2mo ago

Thank you, that's awesome to hear! Curious what the project is, if you can share. And if anything in the API felt rough, an issue would be appreciated.

gemisis 2mo ago
API seems to be fine! I'm admittedly using AI for the integration, so minimal exposure on my end, but agents seem to be good at navigating the API which is good. I'm using it for my project Leviath (leviath.dev), focused on building structured agents. In particular, the way that agents are constructed is via a graph, and so I have a TUI with this that allows both viewing runs, and creating agents with it now in a terminal nicely :D Previously only the web interface really was viable, now it's all doable via CLI which is nice for folks who just want to go through there.
- furkankly 2mo ago
  That seems like a nice fit for rataflow and I'm really happy that it worked well for you! I didn't get to try it out, but Leviath seems like a really cool take on agent orchestration and context management. One question, afaiu, a single agent is composed of multiple stages where each one can possibly make an llm call to a different provider. Does that out-of-the-box multi-stage agent experience lead to the shared append log kind of context window given an agent? I also wonder how this compares to dynamic workflows from Claude Code. Thats one solution to the context bloat that I really love using.
  - gemisis 2mo ago
    > Does that out-of-the-box multi-stage agent experience lead to the shared append log kind of context window given an agent? Leviath supports doing transformations on the context window between stages (https://leviath.dev/docs/stable/stages#carrying-context-across-an-edge https://leviath.dev/docs/stable/stages#carrying-context-acro... has documentation on built in transformations, and scripting custom ones). so you could do a direct transform and then just append onto the region, or you could do something more complex like turn a list of items into a map of key-value pairs on some portion of the list. > I also wonder how this compares to dynamic workflows from Claude Code. I played around with these a bit, and what I found is that I just couldn't get the level of control I wanted. In most instances, I want to design a graph for the agent, and then have that graph be the execution for it (which is exactly where Rataflow comes in great :D https://x.com/GEMISIS/status/2090212911595872556 https://x.com/GEMISIS/status/2090212911595872556) The other thing is that dynamic workflows just seem very.... bloated. Sometimes I want a fairly simple flow, and workflows are just very heavy (whereas for heavier workflows, I haven't found them to give good results)