Your agent passed 10,000 evals. It can still deadlock.

A workflow agent with 20 routing variables has 2^20 or 1,048,576 potential evaluation combinations. Running 10,000 evaluations only samples less than 1% of these possibilities (0.95% precisely). The remaining 99% of combinations are where failures, or "stalls," can occur. A green evaluation suite indicates the workflow can function, but it provides no information about the unseen combinations, including scenarios where all outgoing edges from a node are simultaneously closed, leaving the agent with no path forward.

This failure mode can manifest in research agents with a critic loop: a planner calls a tool executor, a critic node assesses the outcome, and two edges depart from the critic. One edge loops back for further attempts if the retry count is under 3, while the other exits. If the copy-paste error occurs, both edges close, causing the agent to become stuck, burning tokens and wall-clock time without crashing or erroring.

This issue stems from the interaction of multiple guards, which sampling tests are ill-equipped to detect. Tracing tools can illustrate the stalled run but only after the agent has expended resources, potentially retrying against paid APIs multiple times. Observability can only identify stalls that have already transpired and cannot predict whether a stall is possible in an untested graph.

The solution lies in counting the trap states before deployment. The SharpLogic tool takes a workflow graph, converts the guard conditions into a boolean variable set, encodes the graph as a CNF formula (stall formula), and uses zero-dependency Rust to count the satisfying assignments. This count represents the number of condition combinations that result in a workflow stall.

The tool also provides witness assignments that demonstrate how the stall occurs, enabling developers to locate the problematic node and understand the cause. By integrating this counting process into a build pipeline, developers can verify the absence of stalls during development, ensuring the agent functions as intended before deployment.

SharpLogic can handle up to 64 guard variables, with most practical graphs far below this limit. The tool operates in microseconds to sub-millisecond time for real-world workflow graphs, significantly faster than the exponential time required for dense random formulas. SharpLogic is not a proof of the exponential time hypothesis but an engineering solution that operates independently of that theoretical framework.

Developers can try the tool without signing up or installing anything by pasting a graph JSON into the browser trial at sharplogic.vercel.app/#try.