Overview of Your phone is AI’s newest hardware
In this episode of the Stack Overflow Podcast, Ryan Donovan speaks with Div, founder of AGI Inc., about building on-device AI agents for mobile phones that can interact with apps like a human would. The conversation covers how the team optimizes models for mobile chips, trains agents on real devices, evaluates reliability and safety, and envisions a future where phones — and eventually wearables — become the main interface for AI-driven personal computing.
What the company is building
Mobile agents that act like humans
- The core idea is to let an AI operate a phone directly by:
- tapping
- typing
- scrolling
- navigating apps
- The agent is designed to work across third-party apps such as:
- Uber
- Amazon
- DoorDash
- Starbucks
- The goal is to make the phone powerful enough that it can replace much of what people currently do on a desktop.
On-device, not cloud-first
- The company focuses on running models locally on mobile hardware rather than relying entirely on the cloud.
- This requires specialized engineering because mobile chips differ from GPU-heavy cloud infrastructure.
How the technology works
Model architecture and optimization
- The system is described as transformer-based.
- The team uses:
- quantization to shrink models
- compilation and kernel-level optimization for specific mobile processors
- A major challenge is that mobile chips, often from Qualcomm and others, don’t use the same stack as NVIDIA GPUs and CUDA-based cloud systems.
Auto-compilation and self-optimization
- The company built an auto-compilation engine that helps generate optimized code for a given mobile chip.
- This is meant to improve:
- speed
- memory efficiency
- responsiveness
- reliability
Real phones over emulators
- The team initially used emulators, but found them too limited and not representative enough.
- They transitioned to real phones and now run agents on an internal fleet of devices so the system can learn from real-world behavior.
Safety, trust, and guardrails
Human-in-the-loop design
- A recurring theme is making the agent useful without removing user control.
- For risky actions, the AI should:
- show options
- ask for confirmation
- stop before checkout or payment
- hand control back to the user when needed
Restricted and risky use cases
- The company explicitly avoids allowing the AI to use banking apps for financial transactions.
- They also discussed gray areas like:
- dating apps
- screenshots
- spam-like behavior
- The team wants the AI to refuse malicious or risky tasks instead of blindly following instructions.
Defending against attacks
- They simulate scenarios like prompt injection / poisoning attacks, such as hidden instructions inside emails or apps.
- This helps train the agent to detect suspicious behavior and improve its safety model.
Testing and benchmarking
Internal sandboxing
- They built a sandbox environment called yearlywealth.xyz, which clones popular websites for agent testing.
- The cloned sites included examples like:
- DoorDash
- Amazon
- United
- One clone was so realistic that United reportedly sent a legal notice to take it down.
Performance claims
- The company says its agents are highly reliable and cites 97.4% accuracy on Android World, a recognized benchmark for Android-based agents.
- They emphasize that this performance came from their deep device-level optimization work.
Use cases people actually want
Common tasks
Users are already using the agent for:
- ordering coffee
- booking Ubers
- shopping
- entertainment
- daily routines
“Skills” and personal automations
- Users can create custom routines or “skills,” such as:
- “Call me an Uber to the office every morning”
- “When I order food, account for my peanut allergy”
- This turns the product into a personalized assistant rather than just a generic automation tool.
Surprising behaviors
- The team was surprised to see users:
- playing video games with the agent
- finding creative ways to chain tasks together
- This reinforced the idea that people quickly discover new workflows once the system exists.
Scaling challenges and global adoption
Language and regional expansion
- A major next step is supporting more languages and local apps.
- The team noted unexpected viral growth in Thailand, where users began using the product in Thai with local apps.
- They also observed that in many regions — including India and parts of Africa — people are mobile-first or mobile-only, making phones the natural home for this kind of AI.
Different geographies, different behaviors
- Shopping, food ordering, and entertainment were described as the most universal mobile AI use cases.
- But app ecosystems and user expectations vary by region, so the product has to become more localized and multilingual.
Bigger vision
Phone as the new AI hardware
- Div’s long-term view is that the phone may become the first truly universal device for agentic AI.
- The next step could be AI running on:
- smart glasses
- bands
- other wearables
- The broader vision is a hands-free, voice-driven interface similar to a “Jarvis” assistant from Iron Man.
Beyond the phone
- The episode ends with the idea that this could eventually become a new kind of operating system — not just an app.
- A notable endorsement mentioned in the conversation was Rich Miner, a founder of Android, joining as an advisor and suggesting the platform could evolve beyond Android-like paradigms.
Key takeaways
- Mobile devices are becoming the next major frontier for AI agents.
- Building useful on-device agents requires deep work in compilation, quantization, chip optimization, and real-device testing.
- Safety and human control are central, especially for commerce, banking, screenshots, and other sensitive tasks.
- The most successful use cases are the ones that save time on everyday phone actions: shopping, food, rides, media, and routine automation.
- The long-term ambition is a world where you can talk to your phone — or wearable — and it does the rest.