Don't Let LLMs Drive: GraphBit Hard-Wires Agent Workflows with DAGs and a Rust Engine

Forum topic · 小凯 · 2026-05-16

Summary

Most mainstream AI Agent frameworks such as AutoGen, CrewAI, and LangGraph rely on prompt-driven orchestration, where the large language model decides the next step at runtime. This often leads to hallucinated routing, infinite loops, and non-reproducible runs. GraphBit, a research project by Yeahia Sarker, Md Rahmat Ullah, Musa Molla, and Shafiq Joty (arXiv:2605.13848), takes the opposite approach: workflows are pre-defined as DAGs and executed deterministically by a Rust engine. Agents become typed functions with declared inputs and outputs; routing, state transitions, and tool calls are handled by the engine rather than the LLM. A three-layer memory architecture (scratchpad, state, external connectors) isolates context to prevent reasoning degradation in long pipelines. On the GAIA benchmark, GraphBit reportedly achieves 67.6% accuracy—the highest among six frameworks tested—with zero framework-induced hallucinations and only 11.9ms overhead, outperforming AutoGen, CrewAI, and LangGraph. The forum author notes, however, that the gap between GAIA tasks and real production environments should not be overlooked.

Today's AI Agent frameworks mostly let the large language model decide the next step on its own—which is why agents often get lost, loop around, and produce non-reproducible runs. GraphBit's answer: don't let the model drive. Weld the rails in place with DAGs and a Rust engine.

Mainstream frameworks (AutoGen, CrewAI, LangGraph) all use a "prompt orchestration" model—the model decides the workflow itself. This suffers from three chronic problems: hallucinated routing, infinite loops, and non-reproducibility.

GraphBit's approach: define the workflow in advance as a DAG and execute it rigidly with a Rust engine.

Core design

The workflow is fixed before execution. Agents are no longer "thinkers" but "typed functions"—input/output types and what each agent does are written explicitly in the DAG definition.
The Rust engine handles orchestration: routing, state transitions, and tool calls bypass the LLM entirely and are managed by the engine.
A three-layer memory architecture (scratchpad, state, external connectors) isolates context, preventing reasoning degradation in long pipelines.

GAIA benchmark results

Accuracy: 67.6% (highest among tested frameworks)
Zero framework-induced hallucinations
Only 11.9ms overhead
Outperforms six frameworks including AutoGen, CrewAI, and LangGraph

> GAIA is a benchmark for AI agents completing real-world tasks; 67.6% beats all frameworks. But in my view, the gap between GAIA tasks and real production environments should not be ignored.

Paper information

Title: GraphBit: A Graph-based Agentic Framework for Non-Linear Agent Orchestration
Authors: Yeahia Sarker, Md Rahmat Ullah, Musa Molla, Shafiq Joty
Preprint: arXiv:2605.13848 (cs.AI)
Core contribution: DAG + Rust engine agent orchestration; 67.6% on GAIA; zero hallucinated routing
Paper link: https://arxiv.org/abs/2605.13848

References

1. Sarker, Y., et al. (2026). GraphBit. arXiv:2605.13848. 2. Wu, Q., et al. (2023). AutoGen. 3. Hong, S., et al. (2024). MetaGPT.