The problem
Agentic AI plans at the moment it runs. That makes its behavior sensitive to how a prompt is worded and difficult to audit afterwards, which is a poor fit for work where a missed step means a patient is missed.
The arithmetic is unforgiving. Under an idealised model where each step is independent, a ten-step process with 90% per-step reliability completes successfully only 35% of the time. To reach 90% end to end, every step has to be about 99% reliable. Existing workflow platforms give you durable execution and observability, but few guarantees about whether the work was semantically correct.
The approach
GraphFlow treats the workflow diagram as the executable specification: one artifact that defines the data in scope, the execution semantics, and the monitoring.
At compile time, a restricted class of diagrams produces reusable automations whose contracts, meaning their preconditions, postconditions and composition obligations, are intended to be proof-checked before they are admitted to a shared library. At runtime, a durable engine records outcomes in an append-only event log and can enforce those contracts at system boundaries, which is what makes replay, retries and audit possible.
Swimlanes make the trust boundaries explicit, separating verified logic from external systems, human judgment and AI decisions.
What the pilot showed
A year-long pilot across three clinical sites executed 8,728 cohort-enrolled workflow runs with a 97.08% success rate. Observed failures were localised primarily to external integrations rather than to the workflow logic itself.
That result came from an early prototype running without the verified-core subsystem. The formal semantics and the proof-checked admission model described in the paper are specified and under active development, and evaluation of that core is reserved for future work.
Read the paper
The full paper is on arXiv as 2605.14968 under cs.AI, submitted 14 May 2026.