HABIT: A Self-Evolving AI-Agent Harness for GUI-Based CAE Software---Application to Aerospace
This paper introduces HABIT, a self-evolving, host-agnostic AI agent harness designed to automate complex GUI-based CAE workflows in aerospace by synchronizing reliable tool calling with real-time software states and continuously adapting to user habits through a memory-driven learning system.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a world where your computer programs are like incredibly powerful, but slightly stubborn, robots. You tell them what to do, and they try their best. But sometimes, these robots get confused by complex menus, forget what you asked them to do five minutes ago, or accidentally break the very thing they are trying to fix. This is the daily reality for aerospace engineers who use "CAE software"—specialized tools that help them design planes and rockets by simulating wind, stress, and heat. These tools are powerful, but talking to them is like trying to give directions to a robot that only speaks in a very strict, technical language and has a very short memory.
Enter the "AI Agent." Think of this as a super-smart digital assistant that can read your natural language (like "show me the pressure on the wing") and click the right buttons for you. But here's the catch: most AI assistants are like tourists who visit a city once, take a few photos, and then leave. They don't remember your favorite coffee shop, they don't know you prefer your coffee black, and they certainly don't learn how you like to navigate the city over time. For engineers who use these tools every single day, a "one-and-done" assistant isn't enough. They need a partner that evolves, learns their habits, and gets better the more they work together. This is the challenge that a new piece of research aims to solve.
Meet HABIT: The AI That Learns Your Rhythm
A team of researchers from COMAC (the Commercial Aircraft Corporation of China) has built a new kind of AI assistant called HABIT. The name stands for "Harness with Adaptive Behavior Inferred from Trajectories," which is a fancy way of saying: "A tool that watches how you work, learns your favorite shortcuts, and gets smarter every time you use it."
Imagine you have a very talented but forgetful intern. If you ask them to "check the wing," they might check the wrong part, or they might forget to set the colors to your favorite blue. HABIT is different. It's like an intern who not only follows your orders but also keeps a secret notebook of your habits. If you always look at the "WALL" parts of the plane first, or if you always use a specific color map to see heat, HABIT remembers this. Next time you say "analyze this," it doesn't just do the job; it does it your way, automatically.
The Problem with "Dumb" AI
The paper explains that existing AI tools for engineering are like a person trying to drive a car while wearing blindfolds. They might claim they turned the steering wheel, but the car never actually moves. In the world of engineering software, it's not enough for an AI to say it changed a setting; the software actually has to change. If the AI says, "I set the pressure to 500," but the screen still shows 100, that's a disaster.
Furthermore, most AI agents are "stateless," meaning they have no memory of past conversations. They treat every new request as if it's the first time they've ever met you. The researchers argue that for complex engineering work, this is a dealbreaker. You need an agent that knows you, remembers your preferences, and can handle the tricky, multi-step dance of engineering software without getting lost.
How HABIT Works: The Five-Layer Brain
HABIT isn't just a chatbot; it's a sophisticated system built on five layers, designed to be safe, smart, and self-improving.
- The "No-Bluff" Rule: One of the biggest rules of HABIT is that it never lies. If an AI tries to pretend it did something it didn't, HABIT catches it. The system checks the actual software state after every action. If the AI says, "I turned on the red light," but the light is still off, the system knows something went wrong and tries again. It's like a teacher who doesn't just accept "I did my homework" as an answer; they actually check the notebook.
- The Memory Notebook (Passive Learning): This is the coolest part. HABIT has a "memory layer" that passively learns your habits. You don't have to tell it, "Remember this!" Instead, it watches what you do. If you repeatedly choose a specific color map or look at a specific part of the plane, HABIT notices the pattern. After a few sessions, it starts doing those things automatically. It's like a barista who knows you order a latte with oat milk before you even walk in the door.
- The Safety Guardrails: Engineering software is dangerous if used wrong. HABIT has strict rules to keep things safe. For example, it knows that some actions can't be undone (like deleting a file), so it asks for your permission first. It also makes sure that the AI doesn't try to do two things at once that might crash the program. It's like a co-pilot who knows exactly when to take the controls and when to let the human pilot decide.
- The "Context" Trick: To keep the AI from getting confused by too much information, HABIT uses a clever trick called "context engineering." It gives the AI a tiny, compact summary of what's happening on the screen right now, rather than dumping the whole history of the universe into its brain. This keeps the AI focused and fast.
- The Four Modes of Control: HABIT knows when to take charge and when to wait. It has four modes:
- Auto-Pilot: For simple, safe tasks, it just does them.
- Sync Mode: It updates the screen for you as it works.
- Ask First: For big, risky changes, it stops and asks, "Are you sure?"
- Suggestion Mode: It says, "Hey, I think you might want to do this," and you just click "Yes."
The Proof: Does It Actually Work?
The researchers didn't just build HABIT and hope for the best; they put it to the test. They created a suite of 33 different tasks, ranging from simple one-step clicks to complex, multi-step engineering problems. They tested it on real aerospace data, including a famous wing-body model called the CRM-WBT.
The results were impressive. When the AI had to actually change the software and get the final result right (which is the hard part), it succeeded 96.6% of the time. That's a huge jump compared to older methods that just tried to guess the answer in one go. Even better, the AI never "pretended" to do something it didn't do. It was honest, reliable, and consistent.
In one demonstration, an engineer asked HABIT to "analyze the current dataset per my habits." Because HABIT had been watching the engineer over previous sessions, it immediately knew to show only the "WALL" parts of the plane, use a specific "coolwarm" color map, and look at specific slices of the wing. The engineer didn't have to repeat any of those instructions. The AI just knew.
Why This Matters
The paper suggests that this kind of "self-evolving" AI is the future of engineering. It turns the computer from a tool you have to constantly micromanage into a partner that grows with you. Instead of spending hours clicking through menus and setting up the same views every day, engineers can focus on the actual science and design.
HABIT shows that AI doesn't have to be a generic robot that gives the same answer to everyone. It can be a personalized assistant that learns your style, respects your safety rules, and gets better the more you use it. For the aerospace industry, where precision and efficiency are everything, this isn't just a nice-to-have; it's a game-changer. The researchers are already thinking about how to expand this to other types of engineering, like checking how planes handle heat or stress, proving that the sky might not be the limit for these learning machines.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.