Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

QA Engineers Tire of AI Agents 22% More Than Developers: 66.2 vs 54.5 Points

Agents are supposed to take load off engineers, but by evening many feel more tired than before. Habr authors surveyed 26 employees using the NASA-TLX scale, where people rate their own workload. QA averaged 66.2 and developers 54.5, a gap of 22%. An agent writes code in minutes, but the tester answers for a result they never saw being built, and pays for it in stress. When you roll out agents, budget time for verification.

AI-processed from Habr AI; edited by Hamidun News
QA Engineers Tire of AI Agents 22% More Than Developers: 66.2 vs 54.5 Points
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Mental demand is not higher for QA72 vs 77 for developers
Key metric #2QA reports higher effort, time pressure and frustration
Key metric #3Experience with agents and strong models such as Claude Opus reduce the load

AI agents are supposed to take load off engineers: they handle the routine, and tasks close faster. But many people notice the opposite: time is saved, yet they feel more tired by evening than before. Picture a tester to whom an agent hands a large code change in twenty minutes. Writing it by hand would have taken a long time, but now this person has to vouch for its quality, even though they did not see how the agent arrived at the solution. Two Habr authors checked how this kind of work affects their colleagues' well-being and found an interesting pattern. It should help team leads planning agent rollouts, and engineers who want to understand why they feel tired.

How the survey was run

The study covered 26 employees of one company, whom the authors called AI-native: developers, QA and project managers. All of them use a set of tools and specialized agents with confidence and run long multi-step scenarios. In these scenarios, an agent plans actions, writes and edits code, debugs, and calls other tools. Workload was measured with the NASA-TLX method, created at the NASA Ames Research Center. Respondents rate six parameters themselves: mental demand, physical demand, time pressure, effort, their own performance and frustration.

The initial hypothesis was that an IT worker's brain is heavily overloaded when working with agents. It was not confirmed. The average weighted index came out moderately high: about 57 points, with a median of 58 and a range from 25 to 86. A different pattern emerged instead. QA's average index is 66.2, while developers' is 54.5. The difference is 11.7 points, or 22%. If you compare a random tester with a random developer, the tester's workload is higher in about 70% of pairs.

Why testers get more tired

QA's "mental demand" score is lower than developers': 72 vs 77. The gap comes from effort, pace pressure and frustration. The authors explain it this way: generating output has become cheaper, while checking it is still expensive. A tester is responsible for quality and worries about possible damage to their reputation. At the same time, they have less control over how the agent arrived at the solution. Responsibility remains, controllability drops, and that causes stress.

A few more figures for the whole sample. High mental demand (80 and above) was reported by 12 of 26 people, high effort (80 and above) by 11, high frustration (60 and above) also by 12, and strong deadline pressure (80 and above) by 8. Many respondents report growing mental load and irritation even when they save more time overall.

What raises or lowers the load, according to the interviews:

  • Experience with agents. People who are better at steering the process and reducing uncertainty and hallucinations are more confident in the result and worry less.
  • Project type. Vibe-coding a POC or MVP for fun is less taxing than complex production projects with strict requirements for code and security.
  • External pressure. One-day deadlines and mandates like "with AI we should work seven times faster" hit harder than working with the agent itself.
  • Model quality. Strong paid models such as Claude Opus are easier to work with than local MiniMax or Qwen, which freeze, loop and misunderstand the task.
  • Type of work. Development tasks are easier to break into sequential steps. Testing involves more context switching, implicit conditions and reliance on intuition.

What theory says and what to do

The authors draw on Sweller's cognitive load model (1988): intrinsic load reflects the complexity of the task itself, extraneous load comes from interference, and germane load describes the effort of building schemas and gaining deep understanding. AI removes part of the extraneous load: autocomplete, boilerplate code, quick documentation search. The intrinsic load shifts: the question "how do I write this" becomes "how do I ask correctly and verify."

Germane load may fall, and then the ability to read code without hints and to hold a picture of the whole system in your head weakens. Separately, the authors describe "cognitive debt": responsibility for a result that you did not formally create. The brain perceives loss of control as a threat, and anxiety and accumulated fatigue grow.

The article's recommendations are as follows: use AI deliberately, meaning ask it to explain its solution, find weak spots and suggest alternatives; periodically solve tasks without agents; and learn to configure agents systematically, set constraints and manage context.

The authors themselves caution that the sample is small, comes from one company and is not randomized, and the ratings are subjective. They did not originally plan to split respondents into QA and developers, so the conclusion about testers remains an observation, not a proven effect. The practical takeaway is still clear: if you speed up code generation, decide in advance who will verify the result and how. Otherwise the time gain will turn into stress for those who answer for quality. In the second part of the study, the authors promise to examine the effect of AI on engineers' cognitive abilities.

Story

Key Questions Answered

How the survey was run?

The study covered 26 employees of one company, whom the authors called AI-native: developers, QA and project managers. All of them use a set of tools and specialized agents with confidence and run long multi-step scenarios. In these scenarios, an agent plans actions, writes and edits code, debugs, and calls other tools. Workload was measured with the NASA-TLX method, created at the NASA Ames Research Center.

Why testers get more tired?

QA's "mental demand" score is lower than developers': 72 vs 77. The gap comes from effort, pace pressure and frustration. The authors explain it this way: generating output has become cheaper, while checking it is still expensive. A tester is responsible for quality and worries about possible damage to their reputation. At the same time, they have less control over how the agent arrived at the solution. Responsibility remains, controllability drops, and that causes stress. A few more figures for the whole sample.

What theory says and what to do?

The authors draw on Sweller's cognitive load model (1988): intrinsic load reflects the complexity of the task itself, extraneous load comes from interference, and germane load describes the effort of building schemas and gaining deep understanding. AI removes part of the extraneous load: autocomplete, boilerplate code, quick documentation search.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…