#051 • 16 December 2025

Contents of this Newsletter

📝

Why Scripted Testing Breaks As Systems Become Agentic?

⏰ : 4 min

✅

Quash Tests Your Apps

⏰ : 4 min

🔔

What We Shipped Last Week

⏰ : 2 min

📓

Blog Picks

⏰ : 10 min

🗞️

Weekly AI News

⏰ : 5 min

MAIN STORY

Scripts Can’t Keep Up With Agentic Systems

As software systems become more agentic, testing them with static scripts stops making sense. The challenge is not the AI itself, but how AI changes the behavior of mobile apps in ways that are difficult to predict, validate and reproduce with traditional testing approaches.

Let's talk about the distinction between testing AI and testing mobile apps that behave unpredictably because AI is involved. This distinction sounds subtle, but it changes everything.

Right now, a lot of conversations around “AI testing” assume the problem is validating model outputs, checking hallucinations, or evaluating prompts. That matters if you are shipping AI features. But for most mobile teams, that is not the problem they are facing.

They are building mobile apps where AI sits somewhere in the flow and suddenly the app no longer behaves deterministically. With the UI still existing and the APIs still firing, but the path through the app shifts depending on context, state, timing and decisions made upstream by an AI system.

What breaks is predictability rather than intelligence.

Why Agentic Behavior Makes Mobile Testing Harder Than It Looks

Agentic systems change mobile testing because they introduce long-running, stateful behavior into environments that were already fragile. Mobile apps already operate under constraints such as intermittent networks, background execution limits, OS-level interruptions, device-specific quirks and asynchronous UI updates. Adding agentic decision-making on top of this compounds complexity rather than replacing it.

A single user action can now trigger a chain of events that includes UI transitions, backend calls, retries, fallbacks and delayed updates that may surface minutes later. In a checkout flow or a background sync process, a UI success signal may appear immediately while the backend state diverges silently.

From a testing perspective, this breaks several long-held assumptions. Success at the UI level no longer guarantees correctness, because a screen can render correctly while data is partially updated or inconsistent across sessions. Failures also tend to emerge only after multiple steps, once enough state has accumulated for behavior to drift. Reproducing these issues becomes harder because the same flow may not execute the same way twice, even on the same device.

This is where most mobile testing setups start to show cracks. Scripted test cases assume fixed paths, assertions assume immediate outcomes and logs are often disconnected from UI actions. When agentic behavior enters the system, these assumptions collapse quietly rather than failing loudly.

Recent research on agentic system architectures supports this reality. The paper Architectures for Building Agentic AI Systems argues that reliability failures emerge at the system level rather than the model level. It emphasizes the role of planners, execution loops, memory and verification mechanisms in shaping behavior over time instead of validating single outputs. While the paper is not mobile-specific, the implications are especially sharp for mobile environments where state and timing already work against you.

What Needs To Change In How Mobile Apps Are Tested

If mobile apps are becoming more agentic in behavior, testing needs to shift from validating isolated outcomes to validating behavior across execution.

This means observing how UI actions translate into backend effects rather than stopping at whether a button click succeeds. It means validating that intent is preserved across multiple steps instead of only checking that a flow completes. It also means treating logs, API calls and state changes as first-class signals rather than debugging artifacts inspected only after something breaks.

At Quash, this is why we think about testing as execution rather than scripting. Using AI agents to interact with mobile apps allows tests to adapt to changing UI states instead of failing on brittle selectors. Backend validations help ensure that UI success aligns with actual system correctness, while captured execution context makes failures explainable instead of opaque.

The goal is to test mobile apps honestly in a world where behavior is no longer linear.

Agentic systems are not going away and mobile apps will continue to absorb their complexity whether teams plan for it or not. The teams that adapt fastest will be the ones who stop asking whether AI is working and start asking whether their systems behave correctly when intelligence, state and mobile constraints collide.

Your Apps, We Test

Each week, we put our execution engine to work on real apps.

This week, we put our execution engine to work on The Souled Store, using test data to validate the app’s address management flow. Before running the test, we set up an address database containing multiple saved address entries. Using a natural language prompt, we asked Quash to open the app, navigate to the cart, and add a new address. Instead of hardcoding values, Quash used a callable test data action triggered using ‘/’ to select a specific address entry (address_database:2) at runtime. It filled in the address details, saved it as Home and validated that the new address was added successfully.

Want us to test your app next? Drop us a mail at hello@quashbugs.com

What Did We Ship?

Mahoraga Execution Engine

v1.1.2 | December 16, 2025

This week, we shipped the next iteration of our execution engine Mahoraga. This version focuses on making the engine more intelligent, more deliberate and more transparent during test runs.

Mahoraga now uses a stronger planning layer to break prompts into clearer steps and execute flows with better judgment and stability. Alongside this, test reports have been improved with deeper step-level analysis, making it easier to understand not just what happened during a run, but how and why it happened. These updates make executions more reliable and results easier to interpret especially when testing complex mobile flows end to end.

Check it now here: https://quashbugs.com/

Check out our Recent Blogs

News

This Week In AI

NVIDIA Expands Open-Source AI With SchedMD Buy and Nemotron 3 Models
NVIDIA announced it’s bulked up its open-source AI offerings by acquiring SchedMD, the developer behind the widely used Slurm workload manager, while also launching a new family of open AI models called Nemotron 3. The acquisition keeps Slurm open-source and brings critical HPC scheduling tech under NVIDIA’s umbrella, strengthening its position across large-scale AI infrastructure. Together, the model launch and acquisition show NVIDIA expanding beyond hardware into the open-source infrastructure that powers modern AI systems.

OpenAI Ships GPT-5.2 With Stronger Reasoning and Tool Use
OpenAI rolled out GPT-5.2, bringing noticeable gains in multi-step reasoning, code generation, and structured tool execution. The update focuses less on conversational polish and more on reliability in real workflows, especially for developers and enterprises. It’s a direct response to Google’s recent Gemini advances and reinforces the shift from chat-first models to systems that can plan, act, and execute tasks end to end.

Anthropic and Accenture Partner to Scale Enterprise AI Adoption
Anthropic announced a multi-year partnership with Accenture to deploy Claude across large enterprises and regulated industries. Accenture will help organizations operationalize AI with an emphasis on governance, safety, and production readiness. The deal highlights a growing reality: enterprises are choosing partners who can help them run AI systems at scale without breaking trust or compliance.

All Newsletter Issues | Security | Contact Us | Blogs

Why Testing Starts Failing When Apps Become Agentic 🤖

Product

Solutions

By role

By use-case

By industry

Resources

Comparison

Company

Product

Solutions

By role

By use-case

By industry

Resources

Comparison

Company

Get Quash

Product

Solutions

By role

By use-case

Resources

Comparison

Company

Get Quash