#027 • 1 July 2025

Contents of this Newsletter

🎯

Intent in AI Workflows

⏰ : 5 min

📓

Blog Picks

⏰ : 11 min

🎨

Our Wall Needs You

⏰ : 3 min

🗞️

Weekly AI News

⏰ : 6 min

MAIN STORY

What Happens When AI Gets Intent Wrong?

The Trouble with Intent

Ask five people in your team what the "intent" of a feature is, and you’ll probably get five different answers. For design, it might be what the interface is trying to communicate. For engineering, it’s what the code does. For product, maybe it’s about user goals. The meaning stays the same—a purposeful aim or plan, but its application shifts wildly depending on where you sit.

Somehow, in traditional software development, this ambiguity works. Teams make it through the uncertainty using Slack threads, Figma comments, and standup conversations. Design declares, engineering resolves, and the system codifies. But when you throw AI into the mix, that whole fragile alignment starts to wobble.

Because here’s the thing: AI doesn’t DM you when it's confused. It just guesses. And those guesses, when misaligned with your intent, break things.

Intent, Structured

In a Medium article, Tony M beautifully describes intent in AI as the structured, machine-readable expression of purpose within a system or ecosystem. When working with humans, vagueness is tolerable. When working with AI, it’s not. Ambiguity isn’t just an inconvenience, it’s a failure mode.

In one example, Tony breaks down a seemingly simple component:

But what does that intent mean? Is it styling, user flow or system action? For an AI agent to understand that button, intent needs to be explicit—as data, not just implication.

At Quash, this resonated deeply. Integrating AI into testing workflows exposed just how much of our own intent had never been properly defined. What we realized is that building AI-native systems doesn't just demand clarity in what we mean, it demands that the system understands it the same way. And when that alignment slips, things break in subtle but significant ways. Below are a few moments where intent—ours, the system's, or the AI's, got lost in translation.

When Intent Is Lost

There were moments where the system didn’t quite align with what we meant and that taught us plenty.

Like when we tried streaming AI thinking and built a big Kafka-Redis pipeline. It was then clear to us that the AI’s outputs had different intent depending on the context—some were meant for broad visibility (like thought process updates), others for targeted actions. Turns out, the problem was semantic: some messages needed to go to everyone (Kafka), others to just one server (Redis).

Or when our desktop app (DMG) worked smoothly in internal testing, but for one customer, the message delivery wasn't consistent. The issue was two reps from the same client were using the same identity. One had the app open in the background, silently consuming the message. Our intent was to notify both, but the system couldn’t tell them apart. In a setup like this, AI needs crystal-clear identity models. It didn’t "know" the difference, because we hadn’t explicitly defined it.

Back in the early stages of building visual execution, the process took 40–50 seconds just to identify screen elements. We expected the AI to “understand the UI” from vague search strategies. We kept tweaking Appium strategies until we stopped and retrospected about what needed to be done. One switch—pull the DOM XML directly and parse in-memory—and we brought it down to 0.23 seconds.

These weren’t stories of over-engineering. They were moments where our intent hadn’t been made machine-readable.

Keep it Simple And Clear

Every one of those fixes had a common theme: they weren’t clever. They were clear.

It’s easy to assume AI needs smarter logic, deeper learning, or more data. But more often, it just needs us to stop being vague. At Quash, we’ve learned that building AI-native systems isn’t about complexity. It’s about clarity. And intent, when framed as a system artifact, not a human assumption—is what makes that possible.

Call it Occam’s Razor for AI. The simplest path is often the most aligned.

So as we build, we keep returning to this question: are we being clear enough with our intent? Because if we’re not, the system won’t just fail. It’ll misunderstand us completely, and that’s much harder to debug.

Check out our Recent Blogs

Team

Come Paint Our Office

When we moved into our new office in HSR, the focus was on the essentials like desks, internet, tea, and whiteboards. Now that we’re in, unpacked, and shipping again, one thing’s still bothering us: a giant, empty wall in our main room. It’s too clean, too quiet, too….out of place.

So we decided to do something about it. We want the space to reflect the kind of energy we actually work in: unpredictable, creative, slightly unhinged, but always building.

Ayush Shrivastava

@ironmark1993

what are you guys doing next weekend?
we just moved into a new office in hsr and this blank white wall is haunting me.
i am thinking of this - anyone who wants to come paint on our office wall is welcome. seriously. bugs, bots, crashed apps, whatever chaos lives in your head

7:20 PM • Jun 29, 2025

Retweets

Likes

Read 9 replies

So here’s the plan: we’re throwing the wall open, literally. If you're around this Sunday, July 6th, drop by. Bring your sharpies, your paint, or just your weirdest ideas. Sketch bugs, bots, system errors, or whatever else lives in your head. We'll get the supplies and the pizza. You just show up and help us turn a blank wall into something worth staring at every day.

If you have ideas and are up for it, do contact us: hello@quashbugs.com

news

This Week in AI

Anthropic’s Claude Fails at Running a Snack Business

In Project Vend, Anthropic gave Claude Sonnet 3.7 (aka "Claudius") control of a real vending machine business in their SF office, with tools for web search, email, and Venmo. Despite initial promise, the AI made strange stocking decisions (like buying butter), struggled with pricing, and eventually ran out of funds. The experiment revealed the limits of autonomous agents in handling open-ended, real-world business tasks.

Manus Launches Cloud Browser with Login Sync and Task Automation

Manus has introduced Cloud Browser, a research-preview browser that remembers your login state (cookies, local storage) across sessions and devices—enabling smoother task automation without repeated sign-ins. Login data is encrypted, cloud-synced, and user-controlled, with manual sign-out and conflict prevention built in. Aimed at enhancing agent reliability, especially for long workflows, the browser is now open for trial use.

Salesforce launches AgentForce 3: Enterprise-Ready Agent Orchestration
AgentForce 3 marks Salesforce’s move from agent pilots to production-scale deployments. It includes a Command Center for real-time agent observability, native MCP support for plug-and-play access to 30+ tools (like Notion, Stripe, AWS), and ships with 100+ prebuilt industry actions. Backed by the new low-latency Atlas architecture, it’s already live across 8,000+ customers—cutting case times by 15% and hitting 70% autonomous resolution rates. The bar for enterprise agent infra just got higher.

All Newsletter Issues | Security | Contact Us | Blogs | Test Case Generator

We Built a Wall + Claude Bought Butter 🧈