Google Quantum AI has developed a reinforcement learning framework that autonomously controls quantum error correction in real time, removing the reliance on disruptive calibration procedures that halt computation. This approach addresses the critical vulnerability of analogue quantum systems to parameter drift, which degrades performance during long-duration algorithms. By repurposing error-detection syndromes as a continuous learning signal, the system maintains physical gate error rates well below the quantum error correction threshold without interrupting logical operations. The framework employs a multi-objective policy-gradient algorithm that optimizes a surrogate objective function derived from detector event rates. This method circumvents the scalability barriers associated with directly minimizing the logical error rate, which becomes prohibitively expensive to estimate as code distance increases. The algorithm exploits the sparse locality of detecting regions within the circuit, utilizing a factor graph representation to efficiently navigate high-dimensional control spaces. Entropy regularization ensures the agent continues to explore the parameter space, allowing it to adapt to non-stationary environmental drifts. Validation experiments on the Willow quantum processor demonstrated significant performance gains across distance-5 and distance-7 surface codes, as well as a distance-5 colour code. The reinforcement learning agent managed over 1,000 control parameters, steering the processor against injected drift. This real-time control improved logical error rate stability by a factor of 2.4. Integration with decoder steering further enhanced stability by 3.5-fold and delivered a 20 percent additional suppression of the logical error rate beyond the limits of traditional physics-based calibration and human expert tuning. The system achieved record-breaking results, reaching a logical error rate of 7.72(9) x 10^-4 for the distance-7 surface code using the AlphaQubit2 neural network decoder, and 8.19(14) x 10^-3 for the distance-5 colour code with the Tesseract decoder. Scalability projections confirm the framework's applicability to large-scale architectures. Numerical simulations of a distance-15 surface code, encompassing nearly 40,000 control parameters, show exponential convergence rates that are independent of system size. The convergence speed depends on local gate characteristics rather than total qubit count, indicating the method can support the parameter volumes required for practical fault tolerance. The approach requires only error-detection signals and tunable controls, making it broadly applicable to various qubit modalities and error correction codes. This work establishes reinforcement learning as a viable pathway for automating the control of error-corrected quantum systems. By enabling continuous, model-free optimization, the technology offers a robust solution for maintaining fault tolerance in future quantum processors, reducing dependency on manual calibration stacks and supporting uninterrupted execution of complex quantum algorithms.