Can AI agents find water leaks?
A peer-reviewed system coordinated hydraulic simulation, network partitioning, sensor placement and graph-based detection across five water networks. Its strongest result came from simulated leaks equal to 20% of total demand—not a live utility trial or proof that a language model found a physical leak.
The short version
- What happened
- LeakAgent coordinates four specialist components, but the core leak classifier is a dual-branch graph model; the multimodal language model plans and repairs the workflow rather than directly reading pipes.
- Who it affects
- Earlier reliable detection could reduce water loss, disruption and the energy used to treat and pump water.
- What remains unproven
- Leak and normal events were generated in hydraulic simulations; the study did not prospectively detect and confirm physical leaks in a live utility.
Keep in mind
Leak and normal events were generated in hydraulic simulations; the study did not prospectively detect and confirm physical leaks in a live utility.
The headline roughly 96% result used the largest 20% leak magnitude and balanced leak/normal scenarios, so it should not be treated as a universal field accuracy.
Four of five network models were benchmarks; the operational City H model is confidential and cannot be fully reproduced from the public archive.
Noise tests did not cover long-term demand drift, changing pipe roughness, incomplete records, maintenance changes or adversarial sensor manipulation.
The language model coordinates tools; a specialised graph model performs leak detection, and synthetic execution recovery is not a safety evaluation.
Source analysis · 3 linked sources. Primary analysed source date: 6 October 2026.
Read the full report · 8 min The evidence, methods, implications and limitations
The answer is promising in simulation, not proven in live operation
LeakAgent is a coherent attempt to automate a fragmented engineering workflow. It accepts a natural-language request, runs hydraulic simulation, divides a water network into districts, proposes sensor locations and applies a graph model to decide both whether a leak exists and which district contains it. Across five test networks, the authors report strong results when the simulated leak signal was large. That supports further operational testing; it does not show that an autonomous agent has already found leaks in a working city system.
The distinction matters because the headline architecture combines two different kinds of AI. A multimodal large language model coordinates tools, translates instructions and attempts to repair execution failures. The actual leak inference comes from a separately trained Look-Twice Graph Feature Matching model informed by hydraulic pressure sensitivity. Calling the whole system 'agentic' is accurate at the workflow level, but the paper is not evidence that a general-purpose chatbot can inspect sensor streams and reliably diagnose a pipe by itself.[1]
Five networks were tested, but the leak events were generated
The evaluation used EXA5, EXA7, KY3, KY5 and City H, spanning 191 to 920 junction nodes. The first four are public or commercial benchmark models; City H is based on a real operational utility and includes 1,032 pipes, but its location and engineering model remain confidential. The authors provide City H's reported metrics and sensor configuration, while access to the underlying model requires an academic request, non-commercial purpose and confidentiality agreement.
For every network and leak magnitude, the team generated Monte Carlo operating scenarios rather than waiting for observed pipe failures. Eight magnitudes represented 0.5%, 1%, 2%, 3%, 5%, 10%, 15% and 20% of average total demand. Each network-and-magnitude condition was tested with 500, 1,000, 2,000 or 4,000 training samples, balanced one-to-one between leak and normal scenarios. That is a broad simulation matrix, but it cannot reproduce every effect of changing pipe roughness, incomplete asset records, faulty meters, maintenance work or long-term demand drift.[1][2]
The strongest number belongs to the easiest leak condition
At a leak equal to 20% of total demand and 4,000 training scenarios, the paper reports mean global and regional detection accuracy of roughly 96% across the five networks. At 5% magnitude, the model's global accuracy ranged from 87.52% to 95.98% and regional accuracy from 79.62% to 89.92%, depending on the network. At that harder condition it outperformed the tested graph and sequence-model baselines, with the smallest reported advantage being 25.3 percentage points for global detection and 19.6 points for regional detection.
Those comparisons show that the proposed hydraulic features and two-stage graph architecture add information under the authors' benchmark. They do not justify a single universal '96% accurate' label. Accuracy depends heavily on leak size, simulated noise, sensor placement, network structure and class balance. The evaluation uses equal numbers of leak and normal scenarios, whereas real utilities may see a very different prevalence. False-alarm workload and positive predictive value can change sharply when the base rate changes.[1]
Noise tests and ablations help, but field drift remains unmeasured
The researchers varied demand uncertainty and sensor noise up to 5% at a 10% leak magnitude. With 2,000 samples, mean global accuracy stayed near 96% across their three noise scenarios, while regional accuracy remained around 87%. Ablations attributed much of the performance to pressure-sensitivity moments and interactions among graph features. On EXA7, the sensor-placement routine selected 21 sensors across 381 nodes and improved several simulated coverage and detection objectives.
This is useful stress testing within a hydraulic model, but it is narrower than operating a network for months. The authors explicitly identify untested diurnal, weekly and seasonal demand patterns, extreme demand events, changing pipe roughness, boundary-condition uncertainty and incomplete asset records. Sensor placement was optimised with a physics-based detectability surrogate rather than jointly with the final detector. A utility should therefore read the results as an engineering prototype evaluated under controlled uncertainty, not a validated alarm service.[1]
The orchestration results do not establish autonomous safety
The coordinator was compared across 18 proprietary and locally served language models. Task-planning success was reported as 100% on the study's tool-chain tasks, and induced execution failures were usually recovered. Parallel inference took roughly 13 to 32 seconds across the tested networks on one high-end workstation, excluding model training. That suggests the workflow can be operated through a conversational layer without hand-coding each analysis.
Yet synthetic tool errors are not the same as bad operational advice. A coordinator can successfully call the intended function while using stale network data, accepting an invalid assumption or presenting an uncertain result too confidently. The paper does not evaluate cyberattack resilience, malicious instructions, operator over-trust, audit hand-off or the consequences of a false regional diagnosis. For critical infrastructure, the agent interface should remain subordinate to versioned engineering models, access controls, independent alarms and accountable human decisions.[1]
Reproducibility is substantial but incomplete
The authors released code, scenario-generation materials and models for the four benchmark networks, with version 1.1.0 archived on Zenodo. They also provide source data for the main and supplementary figures, including per-repetition detection metrics. That makes the public portion unusually inspectable. The operational City H model cannot be independently rerun from the archive because its utility owner treats the infrastructure details as confidential.
The work was supported by Chinese national and provincial public-research grants, and the authors declare no competing interests. Public funding and code availability reduce some concerns, but an independent utility or laboratory still needs to reproduce the results on a locked release, verify the scenario generator and test networks not selected by the development team. Because water-system topology and sensor quality vary greatly, cross-network transfer in five models is encouraging rather than definitive.[1][2][3]
What would change the assessment
Confidence would rise with a preregistered prospective study in several utilities using concealed, independently curated events: confirmed leaks of different sizes, normal operational disturbances, sensor faults and maintenance changes. It should report time-to-detection, false alarms per network-day, missed leaks, localisation distance, operator workload, repair decisions and water saved—not only balanced-scenario accuracy. A frozen model should be compared with existing utility practice and specialist non-agent baselines.
The decisive evidence would show that performance survives seasonal drift, missing or compromised sensors and incomplete asset records, while operators can understand and override the system safely. Security testing should cover prompt injection and tool misuse in the conversational layer. Until those results exist, LeakAgent is a technically serious, reproducible research system that joins several water-analytics steps. It is not yet evidence that AI agents can be trusted to run leak response unattended.[1][2]