AI operating room scheduling: LLM agent teams reach feasible plans
October 16, 2026
Scheduling an operating room well is hard, and AI operating room scheduling is now being tested as a way to help. A study in npj Digital Medicine asks whether a team of large language model (LLM) agents can write the optimisation code that builds those schedules, and whether a team does better than one model working alone. Across 35 test scenarios and 10 LLMs, the answer was yes, with clear caveats.
How the agent team works
The framework, called iterative multi-agent (IMA), splits the job across four roles: a project manager, an operations research expert, a software engineer and a test engineer. The expert writes the mathematical model. The engineer turns it into Python that runs on the Gurobi solver. The tester checks the output against the scheduling constraints. When a run fails, the project manager reads the error log and sends the task back to the agent that should fix it. The loop stops when a feasible schedule appears or after three attempts.
Multi-agent setups are also being tested beyond scheduling, including in medical exam questions.
Feasibility across models and difficulty levels
A run counted as a success only if the code executed and the schedule met every modelled constraint. Averaged over the scenarios, the IMA framework produced feasible schedules 82.91% of the time. The multi-agent baseline, without the iterative loop, reached 65.59%. A single agent managed 52.53%.
GPT-5.2 was the strongest model. Under IMA it produced a feasible schedule in every scenario, including all 11 hard ones. GPT-3.5-Turbo was the weakest, reaching 36.36% on the hard scenarios.
Retrieval-augmented generation, which gives the agents examples from earlier scheduling work, also helped. In a separate test on three real-world scenarios of 20 cases each, success rose to 100%, 95% and 90% with retrieval, against 60%, 80% and 40% without it. The gain ranged from 15 to 50 percentage points.
Comparison with expert-built models
The authors also compared the generated code with mixed-integer programming models written by operations research experts and solved with the same optimiser. They tested 1,050 synthetic cases and three real datasets: two public sets from Ghent University Hospital in Belgium, and one from West China Hospital. The generated solutions were close to the expert models. The authors report no statistically meaningful cost gap across the West China Hospital departments, although a few scenarios showed small numerical differences.
Hospital case study: utilisation and overtime
The most practical result comes from nine elective surgery departments at West China Hospital, covering 16 weeks. Emergency and ICU cases were excluded. The as-is schedules averaged 84.47% OR utilisation, with 15.60% of scheduled time falling outside regular hours (before 08:00 and after 18:00). The LLM-based schedules averaged 100.98% utilisation and 2.21% overtime. Total OR-days fell from 405.39 to 392.00.
The department totals hide a trade-off. Cardiac surgery needed more OR-days under the LLM schedule, rising from 59.75 to 106.44 per week, while its overtime fell from 48.49% to 3.05%. The authors read the extra days as a more even spread of the workload rather than lost efficiency.
The authors estimate the saving at about USD 22,797 over 16 weeks, using an average operating cost of about USD 1,709 per OR-day, and they call the estimate conservative. Their own inputs (13.39 fewer OR-days at USD 1,709) work out to about USD 22,884, so treat the figure as approximate.
Cost and limits
Running the system with GPT-5.2 cost about USD 0.05 per scheduling instance, based on the token counts the paper reports. The study does not yet cover emergency insertions, surgeon availability or uncertain surgery durations, and it tested nine departments at one hospital. The authors recommend running the system inside hospital-controlled environments, with encryption and de-identification, to protect patient data. They also position it as decision support, with clinicians setting the priorities.
Questions about where people stay in the loop come up in health technology assessment, and the regulatory side is taking shape in the UK framework for AI healthcare regulation.
Source: Tang C, Wei Q, Liu H, Jiang Z, Zhou X, Cao Y, Li K. Solving operating room scheduling problems using large language model based multi-agent architecture. npj Digital Medicine (2026), article in press. doi:10.1038/s41746-026-03319-5
Let Google know we are your trusted source.
Add our editorial as a preferred source in your search results.
Value Science Weekly
The week in health technology assessment, market access and health economics — what moved, and what it means.