IBM introduced the Directed Execution model for the IBM Quantum Compute Service, aiming to give researchers direct control over how quantum circuits are transformed, randomized, and executed before being run on hardware. Rather than keeping error-mitigation logic hidden inside IBM’s servers, the model enables this logic to run on the client side, where it can be inspected, modified, and used as the basis for building custom workflows, without giving up the performance of systems exceeding 100 qubits.
The model is based on a new execution primitive called Executor. It treats the template circuit and its parameters as operations on tensors, allowing large families of circuit instances to be managed without manually assembling them. According to IBM, Executor can run workloads containing 5,000 or more random operations, and it also supports sweeping across noise levels, measurement-basis changes, and parameter sets within a single job using NumPy-like broadcasting rules.
From Hidden Execution to an Inspectable Plan
The Sampler and Estimator primitives created circuit instances on the server to improve performance when handling large experiments, but this made the execution process more like a black box. Directed Execution reintroduces a low-level entry point, while keeping Sampler and Estimator available as high-level interfaces that operate on top of Executor.
The system uses the Boxes and Annotations capabilities introduced in Qiskit SDK 2.0 and 2.1 to identify meaningful regions within a circuit and describe the transformations required on them. The built-in annotations include Twirl for applying Pauli or Clifford twirling, InjectNoise for introducing controllable noise, and ChangeBasis for controlling the rotation of qubits into specified basis states and out of them.
The open-source Samplomatic library then converts the annotated circuit into two outputs: an inspectable template circuit and a structured recipe called Samplex that specifies how randomization and classical post-processing should be performed. Because these two outputs are created locally, researchers can inspect the propagation of transformations before sending the work to the quantum processor.
What Changes for Researchers in Practice?
The error mitigation performed by Sampler and Estimator is now inspectable and customizable instead of being a hidden service. The Qiskit Noise Learning library provides local noise characterization at the circuit-region level, with support for the Pauli-Lindblad and TREX protocols. Qiskit Mitigation includes implementations of methods such as probabilistic error cancellation (PEC), probabilistic error amplification (PEA), TREX, and zero-noise extrapolation using gate folding (ZNE).
The local runtime environment has also become more useful for testing. When the new Estimator is used with a local simulator, the same error-mitigation options can be applied before consuming time on a quantum processing unit. IBM explains that the FakeKingston model simulates a 156-qubit device, a size that is not suitable for full-state simulation locally; it therefore recommends fast approximate checks through Clifford simulation.
IBM presents an example combining PEC, Shaded Lightcones, TREX, and post-selection in a single workflow. According to the results cited in the material, Shaded Lightcones reduced the sampling cost of PEC by approximately 3.4 times while maintaining comparable accuracy within the full noise-model setup.
The Model’s Connection to Quantum Error Correction
Directed Execution is not limited to error mitigation. IBM says that the same local control supports quantum error-correction research, including post-selected quantum error correction and surface-code experiments. It notes that a study on IBM Quantum Nighthawk applied a distance-5 surface code and improved the logical error rate per round by up to 2.8 times by identifying several underperforming components and routing the code around them.
The control extends to the pulse level, including gate timing, rapid qubit resets, and reading soft measurement information before classification. However, these capabilities do not mean that quantum error correction has become a solved problem; they provide a more transparent experimental platform for studying the methods, while the reported results remain tied to specific experiments and tools.
Availability and Limitations
The local Sampler and Estimator primitives were released in version 0.50.0 of qiskit-ibm-runtime, and IBM says that users of the current interfaces will generally need only to update import statements while the body of the code continues to work. The company also released an open-source tool called migrate-qiskit-ibm-runtime to help migrate code from older versions, including V1 and the legacy server-side V2 primitives.
The main value here lies in transferring control and transparency to the researcher, not in launching new quantum hardware. Using Executor directly remains suitable for those who need to build custom workflows or test new protocols, while other users can continue using Sampler and Estimator. The extent to which each experiment benefits from mitigation or correction remains tied to the noise model, sampling cost, and nature of the circuit and hardware used.