Digital Pathology Podcast
Digital Pathology Podcast
194: Medical Agents Fail Real World Stress Tests
This episode is only available to subscribers.
Digital Pathology Podcast +
AI-powered summaries of the newest digital pathology and AI in healthcare papersPaper Discussed in this Episode:
Episode Summary: In this episode, we dive into a comprehensive 2026 benchmarking study that tests whether the highly hyped "Agentic AI" systems are truly ready to revolutionize clinical decision-making. We pit baseline large language models (LLMs) against complex, multi-agent systems in a series of rigorous medical exams and simulated doctor-patient dialogues. The big question: Do the autonomous planning and tool-use capabilities of AI agents actually translate to better diagnostic outcomes, or do they just add unnecessary computational bloat to the clinical workflow?
In This Episode, We Cover:
The Contenders - Baseline LLMs vs. AI Agents: Understanding the difference between a standalone LLM (like GPT-4.1, Qwen-3, or Llama-4) and "Agentic AI" systems (like Manus and OpenManus). Unlike simple chatbots, these agent systems are designed to autonomously reason, plan, and invoke external tools like web browsers, code executors, and text editors to solve complex clinical problems.
The Clinical Gauntlet: How researchers tested these models across three grueling healthcare benchmarks: AgentClinic (step-by-step simulated diagnostic dialogues), MedAgentsBench (a knowledge-intensive medical Q&A dataset), and Humanity’s Last Exam (highly complex, multimodal medical questions designed to defeat AI shortcut cues).
The Verdict - Modest Gains: The surprising reality that despite their advanced, multi-step toolsets, agent systems only yielded a modest accuracy boost over baseline LLMs. We discuss how customized agent models peaked at 60.3% accuracy on AgentClinic MedQA, 30.3% on MedAgentsBench, and struggled at a mere 8.6% on the text-only Humanity's Last Exam.
The Computational Price Tag: Why deploying these agents in a real hospital setting might be completely impractical right now. We discuss the massive inefficiency of these systems, noting that agents like OpenManus consumed more than 10 times the tokens and required more than double the response time compared to a standard LLM.
The Hallucination Problem: Exploring the persistent and dangerous issue of AI "making things up," such as inventing patient statements or assuming test results without asking the patient. We look at how researchers used targeted prompt engineering and an LLM-based output filter to successfully block 89.9% of these clinical hallucinations, though the core problem remains prevalent.
Key Takeaway: While Agentic AI systems show promise by autonomously gathering data and using external tools, their modest accuracy improvements are currently overshadowed by massive computational demands, increased response times, and persistent hallucinations. They represent a step forward in clinical AI architecture, but they remain too inefficient and unrefined for the fast-paced, high-stakes reality of routine clinical deployment.
Uh welcome back trailblazers. You are tuning into a special journal club session of the digital pathology podcast.
Yeah, welcome everyone. It is uh it's great to be back
and today we are well we're cutting through the noise. We aren't talking about what AI might do in 10 years. We are talking about what it's doing right now
right here in 2026 strictly by the evidence.
Exactly. So for this deep dive into the source material, we are pulling apart a really significant new paper that just hit the presses and NPJ digital medicine.
It's um it's a fascinating time to be looking at this specific paper. I mean, we've spent the last few years watching large language models, LLMs, just totally ace medical exams.
Oh, yeah. The USMLA headlines were everywhere,
right? But the question on every clinician's mind and definitely every pathologist's mind has been, sure, they can pass a multiple choice test, but can they actually do the job?
Like, can they handle the messiness of a real patient?
Exactly. The real world ambiguity.
And that is the core mission of the paper we're covering today. It's titled Benchmarking Large Language Modelbased Agent Systems for Clinical Decision Tasks
and the authors Lou Cather and their whole team did an incredible job. It's a massive collaborative effort across multiple institutions
from the National Cancer Center in Beijing all the way to the University of Dresden in Germany. So, it's got a lot of weight behind it.
And trailblazers, this isn't just a literature review. This is a stress test,
a major stress test. We are going to walk you through the method ology, the results, and some rather surprising and frankly maybe a little concerning conclusions.
Yeah, concerning is definitely the right word for some of this data.
So, let's start with the problem statement. The author set the stage pretty clearly. Clinical decision-m is exhausting.
It is incredibly data inensive,
right? And while we've had clinical decision support systems CDSS, for decades now, they were always just a bit rigid,
very rigid. Traditional CDSS or even the early deep learning models were so arrow. They were task specific
like spotting a nodule on an X-ray
or flagging a drug interaction but they couldn't actually reason. Then generative AI comes along GPT4 and its peers
and suddenly we have reasoning.
We do. But the authors point out a really critical flaw here. Accuracy in complex real world simulations where you have dialogue and missing information. It drops significantly.
Yeah. Off a cliff.
Yeah. We are talking about accuracy rates dipping down to the 34. 52% range when things get messy
which is terrifying if you are the patient.
Absolutely.
So the industry shifted we moved from asking a chatbot a question to building agents and this is really the crux of the paper.
It is the shift from zeroot LLMs to aentic AI.
So trailblazers if you take one thing away from this intro it's this shift but let's unpack that because I think a lot of us still use chatbot and agent interchangeable
all the time.
So what is the fundamental difference between me just asking Llama for a medical question and me using an agentic system?
Well, it's a massive architectural shift. When you ask a standard LLM a question, it basically just predicts the next word based on its training.
Like an autocomplete on steroids,
right? It's a zero shot attempt. Think of it like an improv actor. You shout a prompt and they just immediately start talking. No hesitation.
None. They don't stop to think. They don't look anything up. They just go one and done.
Okay. So, it's entirely reactive.
Precisely.
But an agent, specifically the runs tested here, like Manis or open Manis is an autonomous entity. It works in a loop.
A loop.
Yeah. It has a planner, an executive, and a verifier,
right?
When you give it a problem, it doesn't just answer immediately. It pauses.
It actually stops to think.
It does. It thinks to itself, okay, to answer this, I need to check the reference ranges. Then I need to look at the patient history. And maybe I need to run a calculation.
So, it's not just talking. It's thinking about how to think.
Yes, it creates a plan.
And this is crucial. for the study. It can invoke tools
like external tools.
Exactly. The agents in this paper could browse the web. They could execute Python code to run statistics. They could read or edit text files.
Wow.
They function much more like a human resident, gathering data, analyzing it, and then making a decision rather than just blurting out an answer based on memorized patterns. So the hypothesis going into this is that if we give the AI these tools and this agentic freedom,
the ability to Google things and run code,
right? It should absolutely crush the standard dumb LLMs and clinical tasks.
That was the hypothesis. Give it hands and it will build a house essentially. And Lou and Cather set up a very rigorous battleground to test this.
They really did. They didn't just give it a multiple choice test. They used three distinct benchmarks that are honestly pretty clever.
Very clever.
Let's walk the listeners through those arenas because the first one, agent clinic, sounds less like a board exam and more like a video game simulation. really is a simulation. Agent clinic is a role-play environment. They have an AI playing the doctor, which is the agent being tested. Okay?
But they also have an AI playing the patient, another AI playing the moderator, and even a measurement agent that holds the lab results.
So, the doctor agent actually has to interview the patient agent.
Yes. It has to literally ask, "Does it hurt when you breathe?"
And the patient agent responds based on a hidden script.
That is wild.
It is. The doctor has to decide when to order a test, interpret the result, and eventually arrive at a diagnosis and they used cases from the medqa data set and the MIM4V database for this.
Now why is the MIM4V part more important there?
Because MEDQA is textbook stuff. It's clean. MIM4 is based on real electronic health records.
Ah so it's messy.
Very messy, unstructured, full of jargon. It's real life.
Got it. So that's the first arena. Then they had med agents bench.
Right. This is knowledge intensive QA. These are the really hard questions where simple LLMs usually fail. Usually getting under 50% accuracy.
It tests if the agent can synthesize complex medical knowledge.
Exactly.
And then there is the boss level, humanity's last exam, which by the way is just a terrifying name for a test.
It is incredibly dramatic, isn't it? HLE. This is a suite of text and multimodal questions designed specifically to defeat shortcut reasoning.
So, no guessing.
No. If the AI tries to guess based on keywords, it will fail. It requires deep multi-step deduction. It is designed to be the absolute ceiling of current AI capability. Okay, so we had the contenders, the agent systems open Madannis and the proprietary Madness versus the baseline LLMs, Llama 4, GPT4.1, Quinn 3, Gemma 3,
the heavyweights.
Yeah, and I know you trailblazers are waiting for the verdict.
Did the agents with all their web browsers and Python scripts just dominate the field?
This is where it gets nuanced and frankly a bit disappointing for the technoptimists out there. The authors describe the overall games as modest. modest after all that architecture and complexity.
Yeah, let's look at the hard numbers. In the agent clinic using the med QA cases, the textbook ones, the baseline llama 4 model scored 51.4%.
Basically a coin flip,
right? Now, the best performing agent, which was a customized version called open manis medicine tool, scored 60.3%.
Okay, so roughly a 9% bump. That is something. It's not nothing.
It is something. It's a statistically significant improvement over the baseline for that specific data set. But here is A reality check.
Oh,
when they moved to the MIME4 data set, the realworld messy hospital data, the games just vanished
completely
pretty much. Llama 4 was at 24.0%. And the best open manis variant only hit 28.0%.
Wait, 28%.
28%.
That is surprisingly low for both sides. I mean, if a resident was right 28% of the time, they wouldn't be a resident for long.
Exactly. And the difference between the fancy agent and the base IC model wasn't even statistically significant in the MIM SUV tests.
Wow.
It really highlights how hard these realistic simulations are. And when they got to humanity's last exam, the results were humbling across the board.
What were the numbers there?
The baseline LLMs were sitting at about 5 to 8% accuracy. And the highly complex Open Mattis agent, it got 8.6%.
Goodness. So barely a percentage point difference on the hardest tasks.
Correct. So the takeaway from the ACC Accuracy data is yes agents are slightly better. They offer structure
but we are not seeing a revolution where accuracy jumps from 50% to 90%.
It's a grind.
It's a total grind.
And this brings us to the real. So what is this deep dive? Because in healthcare performance isn't just about accuracy. It's about efficiency
resource management.
Right? If I have to wait 10 minutes for a diagnosis or if it costs a fortune to run a single query, that's a massive problem. And this paper did it. deep dive into the cost of doing business with agents.
This is the part that will make hospital administrators sweat. The cost is massive. They measured this in tokens, the pieces of words, the AI processes, and in actual time.
Give us a comparison. How much more expensive is the agent?
Well, in the agent clinic, the baseline llama 4 used about 14,000 tokens per scenario to get to a diagnosis.
That's 14,000.
The open manis agents, they use between 92,000 and 168,000 tokens.
That is a 10-fold increase. more than 10fold in some cases and it translates directly to time. Llama 4 could give you a diagnosis in about 32 seconds.
Very fast. The best open manis agent took nearly 69 seconds and the proprietary manis system was even slower in some context just because of network latency.
So let me get this straight. If I'm a clinician, you're asking me to wait twice as long and pay 10 times the compute cost
for a seven or 8% accuracy boost on clean data.
And in real world data, maybe no boost at all.
Exactly. The authors call this a clear imbalance. For any trailblazer looking to implement this right now, you have to ask if that trade-off is viable.
It doesn't seem like it.
Currently, the study suggests it's a steep price to pay for marginal gains. You're burning a massive amount of energy for a very small step forward.
Now, one thing I found really interesting in the methodology, and this actually gives me a bit of hope for the future, was that they didn't just run the code out of the box,
right? They customized it.
Yeah. They tried to manage the agents. They compared the original Open Manis to something they called Open Manis medicist.
This was a brilliant part of the study. The original agent was a bit chaotic. It would use tools randomly. It would browse the web when it didn't need to.
Like a distracted student.
Exactly. So, the authors applied prompt engineering. They gave the agent a persona. They essentially told it, "You are a physician assistant."
I love that. Act like a PA. And just giving it a job title changed how it worked.
It did. They measured this using a concept called cyclomatic complexity.
Okay, break that down for us. What is cyclatic complexity?
Imagine a ball of headphones in your pocket. Tangled, messy, hard to trace a line from one end to the other.
A nightmare,
right? That's high cyclatic complexity in code or logic. Lots of loops and confusing jumps. Low complexity is like a neatly coiled cable.
Yeah.
The medicist version significantly reduced that complexity. It streamlined the workflow. Instead of jumping all over, the place the agent followed a more logical linear path.
So, it cleaned up the thought process, but did it actually make it smarter?
It made it more consistent. And when they added explicit instructions on how to use tools, which was the medicine tool version, that's where they saw that accuracy bump up to 60.3% in Med QA.
So, it really shows that even robots need good management.
Absolutely. You can't just set an agent loose on the internet. You have to give it a standard operating procedure.
Let's pivot to something that is obviously crucial for pathology and medicine in general.
Multimodal performance.
Yes. Images.
We aren't just reading text. We're looking at images.
Yeah.
So, they tested these systems on cases from the New England Journal of Medicine, NEJM, that included clinical photos and X-rays.
How did the agents fare there?
This was perhaps the most counterintuitive finding in the whole paper. They set up two scenarios. In scenario A, they gave the AI the image right at the start. Here's the patient history. Here is the X-ray
up front.
Up front. And In that case, the agent Vopen Manis did slightly better than the baseline. 27.5% versus 25.0%.
Okay, so that's consistent with the text results. Slight edge to the agent,
right? But in scenario B where the AI actually had to ask for the image simulating a doctor deciding, hm, I should order an X-ray. The baseline llama 4 actually beat the agent system.
Wait, hold on. The simple model beat the complex agent specifically when the agent had to choose to use a tool.
Yes. Llama 4 hit 30.0% and the agent dropped to 29.2%.
Why would having a choice make it worse?
The authors attribute it to something called internal tool competition.
Tool competition like the tools are fighting each other
in a way. Yeah. The agent has all these options. Web search, code execution, image analysis. When it has to make a choice, it sometimes gets confused or prioritizes the wrong thing entirely.
Oh, I see.
It might try to Google the visual symptoms instead of just looking at the because it thinks the web search is a safer bet.
It's a classic case of paralysis by analysis.
Exactly. The baseline model just looked at whatever data was available and made a call.
That is a huge lesson. More complexity doesn't always equal better outcomes if the orchestration isn't perfect. It literally overthought the problem.
Absolutely. It's like having a Swiss Army knife but struggling to pull out the right blade while the other guy just uses a screwdriver and gets the job done.
All right, we have to talk about the elephant in the room.
The safety check.
The thing that keeps hospital lawyers up at night. Hallucinations.
Yeah, this part of the paper is sobering. They tracked how often the agents just made things up, specifically fabricating patient statements or inventing lab results that never happened.
The numbers were high.
They were prevalent. In the med QA data set, the original Open Manis agent hallucinated in 97.2% of the scenarios.
97.2%. That is practically every single time. That means if I use this tool a 100 times, 90 seven times it's going to lie to me about something.
Yes, it would invent a patient response or assume a lab value was normal without actually checking.
That's terrifying.
However, and this is the good news, the study implemented a safeguard. They used a llama 4 postprocessing module. Think of it as an editor at the end of the line.
Did the filter work
surprisingly well? It blocked 89.9% of those hallucinations before they ever became the final output. So, the user never saw them. Okay, that's a relief.
But even with the filter hallucinations still impacted the final diagnosis in about 30% of the scenarios.
So one in three diagnosis was based at least in part on madeup information.
That just sounds like a complete deal breaker.
You would think so.
Yeah.
But here is the surprise twist regarding the data.
Oh.
The authors analyze the accuracy of the cases with hallucinations versus the cases without hallucinations.
I'm guessing the ones with lies were way less accurate.
You'd be wrong. There was no statistically significant difference in diagnostic accuracy between the two.
Yeah. How can you hallucinate medical facts and still get the right answer? That makes no sense.
It's a paradox. The authors Lou and the team theorized that the hallucinations might have actually prompted the agent to gather more information.
How so?
Well, for example, if it hallucinated a weird lab result, that error might have triggered a double check mechanism or a web search to verify it, which ultimately led it down the right path to the correct diagnosis.
So, it stumbled its way to the truth. It basically failed upwards.
In a way, yes, it used the hallucination as a stepping stone. It's a fascinating, if slightly unsettling, emergent behavior. It suggests that the path to the right answer isn't always a straight line of truth for an AI.
But we can't rely on that in a hospital. I mean, your honor, I hallucinated, but I got it right. Is not going to hold up in court.
No, it certainly won't. And relying on serendipitous error is definitely not a safety strategy. So, let's bring this all together for our trailblazers listening. We've looked at the accuracy, the massive token usage, the workflow customization, and this wild hallucination issue. What is the final verdict from the authors?
The synthesis from the paper is clear, but very cautious. Agents offer a necessary structure. They allow for a tool access and planning, which did lead to those small accuracy wins we discussed,
right?
But as of 2026, the technology is currently too slow, too computationally expensive, and just too prone to hallucination for for routine clinical deployment.
It sounds like we're very much in the prototype phase. We have the concept car, but it's not ready for the highway.
Exactly. The efficiency versus accuracy trade-off is the biggest hurdle. A 7% game just isn't enough to justify the resource drain and the latency.
Not at all.
We need these systems to be faster and cheaper before they can sit on a doctor's desktop. And frankly, we need them to stop making things up, even if they eventually get the answer right.
Because trust is the currency of healthcare. If I can't trust trust the data the agent is using. I just can't use the agent
without a doubt.
This has been an incredible breakdown of the reality of AI in 2026. It's not just magic, it's engineering, and it clearly has limits.
And that's exactly why papers like this one in NPJ digital medicine are so important. They strip away the hype and give us the actual benchmarks.
Yeah,
it's not about will AI replace us. It's about is this tool even worth the electricity it consumes right now.
For all of you listening, I highly encourage you to pull up the full paper. Look, with the word flow graphs, seeing how the medicine prompt cleaned up the decision tree. That visual of the cyclatic complexity dropping, it's a literal masterclass in prompt engineering.
Yeah,
it really is visually striking to see the difference.
We will leave you with a final thought to mull over. We talked about how hallucinations didn't drop accuracy because they maybe forced the AI to doublech checkck itself,
right? The paradox.
It makes you wonder, is creativity or error a necessary component of complex reasoning? I mean, in human learning, We learn by making mistakes. Does an AI need the freedom to be wrong in order to figure out how to be right?
That is the ultimate question for the next generation of AI development. If we squash the hallucinations completely, do we also squash the reasoning capability?
Keep thinking about that, Trailblazers. Explore it on your own. Until next time, keep blazing trails.
Chat
🩺
Benchmarking Agentic AI Systems for Clinical Decision Tasks
2 sources
These sources evaluate the performance of agentic AI systems against standard large language models in clinical decision-making tasks during the year 2026. While agent systems utilize specialized architectures to reason, plan, and invoke external tools like web browsers or code executors, they demonstrate only modest accuracy improvements over simpler models. This marginal gain comes at a high cost, requiring over ten times the computational power and significantly longer response times. Research indicates that while task-specific prompt engineering can streamline complex workflows, hallucinations remain a prevalent risk, appearing in the vast majority of simulated scenarios. Furthermore, the studies reveal a "paralysis by analysis" effect where increased tool options can occasionally degrade performance compared to baseline models. Ultimately, the sources conclude that current agentic designs are insufficiently efficient and reliable for routine medical deployment.
How does agentic AI differ from standard LLMs in clinical tasks?
What are the computational costs and response time trade-offs for agents?
How do hallucinations impact the diagnostic accuracy of medical AI agents?
Today • 10:28 AM
can you create show notes for the audio overview and model it after the show notes pasted below with max 3500 characters: Paper Discussed in this Episode:
Artificial Intelligence-Based Digital Image Analysis for Assessing Ki67, P53, and PHH3 Expression in Glioblastoma Multiforme. Devrim T, Erkilinc G, Tuncer SS. J Coll Physicians Surg Pak 2026; 36(02):153-157
Episode Summary: In this journal club deep dive, we step out of the theoretical future of AI and look at a direct, hard-data showdown between artificial intelligence and the human eye. We examine a groundbreaking 2026 study on Glioblastoma Multiforme (GBM) that forces us to ask an uncomfortable question: What happens when the AI and the human completely disagree? And more importantly, is it possible that the AI is detecting a biological reality that experienced human pathologists are entirely missing?
In This Episode, We Cover:
• The "Boss Battle" of Neuro-Oncology: Understanding Glioblastoma Multiforme (GBM), the most aggressive primary brain tumor in adults, and why precise prognosis dictates the entire treatment strategy.
• The Big Three Biomarkers (The Speedometer, The Brakes, and The Neon Sign):
◦ Ki67: The "speedometer" of the tumor, marking active cell proliferation.
◦ p53: The "guardian of the genome," acting as the emergency brakes for damaged cells. In GBM, these brakes are often broken or mutated.
◦ PHH3: A specific "neon mitosis tracker" that lights up dividing cells, offering a cleaner alternative to traditional manual counting.
• The Showdown - Humans vs. AI: Two experienced pathologists go head-to-head with an AI digital image analysis system (QuantCentre module by 3DHISTECH) on 20 adult GBM cases, looking at both 1 mm² and 7 mm² tumor hotspots.
• Round 1 - The Shocking Lack of Concordance: The AI and human pathologists had practically zero statistical agreement (Cohen's Kappa) on the raw numbers. The human eye acts interpretively, filtering out background noise, while the AI calculates literal pixel intensity.
• Round 2 - The AI's "Aha!" Moment: Biologically, a high proliferation rate (Ki67) must correlate with high mitosis (PHH3). Human pathologists failed to find any statistically significant link between these markers. The AI, however, found strong, biologically accurate correlations between Ki67 and PHH3, and between PHH3 and p53.
• The Future of the Lab: Why AI shouldn't replace pathologists, but rather serve as a hyper-sensitive tool to uncover hidden data patterns and personalize medicine. We also discuss the major roadblock preventing immediate clinical rollout: color standardization and image quality.
Key Takeaway: The lack of agreement between humans and machines doesn't mean the AI is wrong. By successfully identifying crucial biological relationships that humans missed due to attentional fatigue and subjectivity, the AI proved its data might actually be closer to the biological truth than our current gold standard.
Paper Discussed in this Episode:
Benchmarking large language model-based agent systems for clinical decision tasks. Liu, Y., Carrero, Z.I., Jiang, X. et al. npj Digit. Med. 2026.
Episode Summary: In this episode, we dive into a comprehensive 2026 benchmarking study that tests whether the highly hyped "Agentic AI" systems are truly ready to revolutionize clinical decision-making. We pit baseline large language models (LLMs) against complex, multi-agent systems in a series of rigorous medical exams and simulated doctor-patient dialogues. The big question: Do the autonomous planning and tool-use capabilities of AI agents actually translate to better diagnostic outcomes, or do they just add unnecessary computational bloat to the clinical workflow?
In This Episode, We Cover:
The Contenders - Baseline LLMs vs. AI Agents: Understanding the difference between a standalone LLM (like GPT-4.1, Qwen-3, or Llama-4) and "Agentic AI" systems (like Manus and OpenManus). Unlike simple chatbots, these agent systems are designed to autonomously reason, plan, and invoke external tools like web browsers, code executors, and text editors to solve complex clinical problems.
The Clinical Gauntlet: How researchers tested these models across three grueling healthcare benchmarks: AgentClinic (step-by-step simulated diagnostic dialogues), MedAgentsBench (a knowledge-intensive medical Q&A dataset), and Humanity’s Last Exam (highly complex, multimodal medical questions designed to defeat AI shortcut cues).
The Verdict - Modest Gains: The surprising reality that despite their advanced, multi-step toolsets, agent systems only yielded a modest accuracy boost over baseline LLMs. We discuss how customized agent models peaked at 60.3% accuracy on AgentClinic MedQA, 30.3% on MedAgentsBench, and struggled at a mere 8.6% on the text-only Humanity's Last Exam.
The Computational Price Tag: Why deploying these agents in a real hospital setting might be completely impractical right now. We discuss the massive inefficiency of these systems, noting that agents like OpenManus consumed more than 10 times the tokens and required more than double the response time compared to a standard LLM.
The Hallucination Problem: Exploring the persistent and dangerous issue of AI "making things up," such as inventing patient statements or assuming test results without asking the patient. We look at how researchers used targeted prompt engineering and an LLM-based output filter to successfully block 89.9% of these clinical hallucinations, though the core problem remains prevalent.
Key Takeaway: While Agentic AI systems show promise by autonomously gathering data and using external tools, their modest accuracy improvements are currently overshadowed by massive computational demands, increased response times, and persistent hallucinations. They represent a step forward in clinical AI architecture, but they remain too inefficient and unrefined for the fast-paced, high-stakes reality of routine clinical deployment.
Studio
Audio Overview
Slide Deck
Video Overview
Mind Map
Reports
Flashcards
Quiz
Infographic
Data Table
Medical Agents Fail Real World Stress TestsDeep Dive · 1 source · 6d ago