Instantiating an agent is not the bottleneck for LLMs. Two hundredths of a second is a rounding error compared to what the model costs in time.","title":null,"type":"comment","url":null}],"created_at":"2025-03-06T23:28:38.000Z","created_at_i":1741303718,"id":43286072,"options":[],"parent_id":43277689,"points":null,"story_id":43274435,"text":"Wrong, actual code and tests provided than show 10000x speed up. Users can run it themselves and have been seeing better results.
Appreciate if you didn\u2019t make up stuff.","title":null,"type":"comment","url":null}],"created_at":"2025-03-06T08:08:06.000Z","created_at_i":1741248486,"id":43277689,"options":[],"parent_id":43276332,"points":null,"story_id":43274435,"text":"They just made it up.","title":null,"type":"comment","url":null}],"created_at":"2025-03-06T04:10:34.000Z","created_at_i":1741234234,"id":43276332,"options":[],"parent_id":43274435,"points":null,"story_id":43274435,"text":"how did you arrive at this number 10000?","title":null,"type":"comment","url":null},{"author":"esafak","children":[],"created_at":"2025-03-06T05:29:51.000Z","created_at_i":1741238991,"id":43276720,"options":[],"parent_id":43274435,"points":null,"story_id":43274435,"text":"LangGraph, not LangChain.","title":null,"type":"comment","url":null},{"author":"tomnipotent","children":[{"author":"AustinDev","children":[{"author":"tough","children":[],"created_at":"2025-03-08T14:53:22.000Z","created_at_i":1741445602,"id":43300717,"options":[],"parent_id":43287585,"points":null,"story_id":43274435,"text":"LangChain has several products and has been building on the Agent space for years
im a fan of vibe coding but that's kinda of a stretch
lmfao","title":null,"type":"comment","url":null}],"created_at":"2025-03-07T05:34:25.000Z","created_at_i":1741325665,"id":43287585,"options":[],"parent_id":43276938,"points":null,"story_id":43274435,"text":"I mean, both companies are just things I could have cursor do in a few hours. So probably not.","title":null,"type":"comment","url":null}],"created_at":"2025-03-06T06:11:38.000Z","created_at_i":1741241498,"id":43276938,"options":[],"parent_id":43274435,"points":null,"story_id":43274435,"text":"This "10,000x" faster claim is specific to how long it takes to instantiate a client object, before actually interacting with it.
Turns out the LangGraph code uses the official OpenAPI library which eagerly instantiates an HTTPS transport and 65% of runtime was dominated by ssl.create_default_context (SSLContext.load_verify_locations) when I tested using pyinstrument. This overhead is further exasperated by the fact that it's happening twice - once for the sync client, and a second time for the async client. Rest of the overhead seems to be Pydantic and setting up the initial state/graph.
Agno wrote their own OpenAPI wrapper and defers setting up the HTTPS transport during agent creation, so that cost still exists it's just not accounted for in this "benchmark". Agno still seems to be slightly faster when you control for this, but amortized over a couple of requests it's not even a rounding error.
I hope the developers get rid of this "claim" and focus on other merits.","title":null,"type":"comment","url":null},{"author":"dcreater","children":[{"author":"torginus","children":[{"author":"mountainriver","children":[],"created_at":"2025-03-07T03:29:35.000Z","created_at_i":1741318175,"id":43287178,"options":[],"parent_id":43277527,"points":null,"story_id":43274435,"text":"Holy cow! Did you say 10000x?!?!","title":null,"type":"comment","url":null}],"created_at":"2025-03-06T07:44:20.000Z","created_at_i":1741247060,"id":43277527,"options":[],"parent_id":43276972,"points":null,"story_id":43274435,"text":"But at least this one is useless 10000x faster!","title":null,"type":"comment","url":null},{"author":"barbazoo","children":[],"created_at":"2025-03-06T17:53:17.000Z","created_at_i":1741283597,"id":43283041,"options":[],"parent_id":43276972,"points":null,"story_id":43274435,"text":"Why is it unnecessary? I'm genuinely interested, it's not like JS where we have a plethora of industry tested frameworks to choose from.","title":null,"type":"comment","url":null}],"created_at":"2025-03-06T06:17:49.000Z","created_at_i":1741241869,"id":43276972,"options":[],"parent_id":43274435,"points":null,"story_id":43274435,"text":"Another day another unnecessary ai framework.","title":null,"type":"comment","url":null},{"author":"yuzhun","children":[],"created_at":"2025-03-06T07:39:52.000Z","created_at_i":1741246792,"id":43277493,"options":[],"parent_id":43274435,"points":null,"story_id":43274435,"text":"Tried agno. Its API has less mental burden than langchain. As for the speed advantage, it hasn't been noticed much.","title":null,"type":"comment","url":null},{"author":"thecleaner","children":[],"created_at":"2025-03-06T08:18:07.000Z","created_at_i":1741249087,"id":43277760,"options":[],"parent_id":43274435,"points":null,"story_id":43274435,"text":"Congratulations on the release. Although I hope the developers take the lesson that AI frameworks are unnecessary. You don't need frameworks to write HTTP calls. Just a good enough SDK would do.","title":null,"type":"comment","url":null},{"author":"moltar","children":[],"created_at":"2025-03-06T11:55:27.000Z","created_at_i":1741262127,"id":43279171,"options":[],"parent_id":43274435,"points":null,"story_id":43274435,"text":"But only in Python.","title":null,"type":"comment","url":null},{"author":"turnsout","children":[{"author":"AStrangeMorrow","children":[],"created_at":"2025-03-07T18:17:04.000Z","created_at_i":1741371424,"id":43292599,"options":[],"parent_id":43281512,"points":null,"story_id":43274435,"text":"I\u2019d wager it can probably a few percents of the full runtime. But no matter the variation in the time it takes for LLMs to generate outputs (depending on the nb of tokens produced/input size) likely drowns it completely.
Like 5s+/-1s vs 4.95s+/-1s","title":null,"type":"comment","url":null}],"created_at":"2025-03-06T15:45:32.000Z","created_at_i":1741275932,"id":43281512,"options":[],"parent_id":43274435,"points":null,"story_id":43274435,"text":"Is Python execution even a rounding error in the full execution time for a LangChain flow?","title":null,"type":"comment","url":null},{"author":"eternityforest","children":[{"author":"bediashpreet","children":[{"author":"eternityforest","children":[],"created_at":"2025-03-08T06:53:20.000Z","created_at_i":1741416800,"id":43298101,"options":[],"parent_id":43286545,"points":null,"story_id":43274435,"text":"7B is pretty big for CPU though","title":null,"type":"comment","url":null}],"created_at":"2025-03-07T00:51:14.000Z","created_at_i":1741308674,"id":43286545,"options":[],"parent_id":43286279,"points":null,"story_id":43274435,"text":"7B is the sweet spot","title":null,"type":"comment","url":null}],"created_at":"2025-03-07T00:00:19.000Z","created_at_i":1741305619,"id":43286279,"options":[],"parent_id":43274435,"points":null,"story_id":43274435,"text":"Can this handle smaller models like Qwen 1.5B, or does it need some real intelligence to get the tool calling to work?","title":null,"type":"comment","url":null},{"author":"slake","children":[],"created_at":"2025-03-08T04:36:38.000Z","created_at_i":1741408598,"id":43297535,"options":[],"parent_id":43274435,"points":null,"story_id":43274435,"text":"Does it work with o1 type reasoning models. I had trouble running phidata (the old named framework) with it.","title":null,"type":"comment","url":null}],"created_at":"2025-03-05T23:54:16.000Z","created_at_i":1741218856,"id":43274435,"options":[],"parent_id":null,"points":47,"story_id":43274435,"text":null,"title":"Agno: Agent framework 10,000x faster than LangChain","type":"story","url":"https://docs.agno.com/introduction"}