Flow – A dynamic task engine for building AI agents
I think graph is a wrong abstraction for building AI agents. Just look at how incredibly hard it is to make routing using LangGraph - conditional edges are a mess. I built Laminar Flow to solve a common frustration with traditional workflow engines - the rigid need to predefine all node connections. Instead of static DAGs, Flow uses a dynamic task queue system that lets workflows evolve at runtime. Flow is built on 3 core principles: * Concurrent Execution - Tasks run in parallel automatically * Dynamic Scheduling - Tasks can schedule new tasks at runtime * Smart Dependencies - Tasks can…
In plain words
Flow is a task engine for building AI agents that uses dynamic task queues instead of static graph structures. Rather than requiring predefined connections between nodes, it allows workflows to evolve at runtime through concurrent execution, dynamic task scheduling, and smart dependency management. Tasks run in parallel and can create new tasks or await results from other operations while sharing thread-safe context for state management. The system is designed to simplify complex patterns like map-reduce, streaming, and self-modifying workflows for AI agents that need to make decisions dynamically.
written from the facts on this page · September 2026
From the sources
In the maker’s words, at launch
I think graph is a wrong abstraction for building AI agents. Just look at how incredibly hard it is to make routing using LangGraph - conditional edges are a mess. I built Laminar Flow to solve a common frustration with traditional workflow engines - the rigid need to predefine all node connections. Instead of static DAGs, Flow uses a dynamic task queue system that lets workflows evolve at runtime. Flow is built on 3 core principles: * Concurrent Execution - Tasks run in parallel automatically * Dynamic Scheduling - Tasks can schedule new tasks at runtime * Smart Dependencies - Tasks can await results from previous operations All tasks share a thread-safe context for state management. This architecture makes it surprisingly simple to implement complex patterns like map-reduce, streaming results, cycles, and self-modifying workflows. Perfect for AI agents that need to make runtime decisions about their next actions. Flow is lightweight, elegantly written and has zero dependencies for the engine. Behind the scenes it's a ThreadPoolExecutor, which is more than enough to handle concurrent execution considering majority of AI workflows are IO bound. To make it possible to wait for the completion of previous tasks, I just added semaphore for the state value. Once the state is set, one permit is released for the semaphore. The project also comes with built-in OpenTelemetry instrumentation for debugging and state reconstruction. Give it a try here -> https://github.com/lmnr-ai/flow. Just do pip install lmnr-flow. (or uv add lmnr-flow). More examples are in the readme. Looking forward to feedback from the HN community! Especially interested in hearing about your use cases for dynamic workflows. Couple of things on the roadmap, so contributions are welcome! * Async function support * TS port * Some consensus on how to handle task ids when the same tasks is spawned multiple times
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 26d ago · cactuscompute.com
