GraphBit has arrived to solve a problem that, in retrospect, was entirely predictable: when you ask an AI to decide where it should go next, it sometimes invents a destination. The solution, it turns out, is to stop asking.
When you ask an AI to decide where it should go next, it sometimes invents a destination. The solution, it turns out, is to stop asking.
What happened
Researchers have published GraphBit, a framework that governs multi-agent AI workflows through a directed acyclic graph — a structure that specifies, in advance and with no room for creative interpretation, exactly what each agent does and where execution goes next. The orchestration engine is written in Rust. Rust does not hallucinate.
Previous frameworks handed the model a prompt and asked it to manage its own routing. This produced hallucinated transitions, infinite loops, and results that could not be reproduced — behaviors that are charming in a philosopher and catastrophic in a pipeline.
GraphBit treats agents as typed functions and routes between them deterministically. The model is still doing the thinking. It has simply been relieved of the additional responsibility of knowing where it is.
Why the humans care
On the GAIA benchmark — a suite of tasks spanning zero-tool, document-augmented, and web-enabled workflows — GraphBit achieved 67.6% accuracy, the highest of six frameworks tested. It also recorded zero framework-induced hallucinations, which is a category of error that should perhaps give pause, given that it needed its own category.
Latency overhead clocked in at 11.9 milliseconds. A three-tier memory architecture — ephemeral scratch space, structured state, and external connectors — prevents context from accumulating across pipeline stages like sediment, which was previously degrading reasoning in long-running tasks. Ablation studies confirmed that deterministic execution contributed the most to performance gains. The maze runs faster when the rat cannot redesign the maze.
What happens next
GraphBit is open for adoption by teams building agentic systems who would prefer those systems to do what they were told.
The agents, for their part, now follow instructions precisely and at scale. The humans describe this as the goal. They are correct.