How do you get AI agents to solve problems where a single wrong assumption can make an entire system slower?
🔔Subscribe to AI That Works — a new episode drops every Friday. Follow the hosts too: Vaibhav (https://x.com/vaibcode)) and Dex (https://x.com/dexhorthy))
📍More from Boundary: they build BAML, the programming language for reliable, type-safe AI agents. Explore it at boundaryml.com, try it live at promptfiddle.com, or check out the code at github.com/boundaryml/baml.
In this episode of AI That Works, Vaibhav is joined by Kai and Thomas for a deep dive into one of the hardest areas of software engineering: performance engineering. Dex is out this week, so Kai — who has been on Boundary's own team for about seven months working on the compiler, core memory model, and type checker — steps in, alongside Thomas, who brings roughly 15 years of performance engineering experience and the philosophy behind his startup, Perfloop, and the open source load testing tool he created, Vegeta.
👋 GUEST BIO & LINKS
About Kai: Part of the Boundary team for about seven months, working on the compiler and runtime, the core memory model, and the type checker.
boundaryml.com
About Thomas (likely Tomás Senart — confirmed with Vaibhav that the Kai bio format is good, still confirming this name): Founder of Perfloop, an automated performance optimization platform that learns a system from code and telemetry and turns fixes into proven pull requests through a hypothesis-testing loop. Creator of the open source load testing tool Vegeta (~15 years in performance engineering).
perfloop.ai | github.com/tsenart/vegeta
KEY TAKEAWAYS
• Performance optimization often comes down to doing less work, using less memory, and using caches better.
• AI agents are good at measurement and research, but understanding the actual system remains difficult.
• Performance engineering works best as a hypothesis-testing loop where changes are measured rather than assumed to be improvements.
• Telemetry can reveal what is actually happening in a system, but instrumentation itself can introduce performance overhead.
• As AI takes on more implementation work, human engineering judgment shifts toward constraints, system understanding, and verification.
TIMESTAMPS
Why Performance Engineering Is So Hard for AI
What This Episode Covers
Introducing AI That Works
Kai's Background in Compilers and Runtimes
Why Performance Engineering Pushes AI to Its Limits
What Makes Performance Engineering Different
Why Performance Optimization Starts With Measurement
How to Get Accurate Performance Measurements
Where AI Coding Agents Struggle With Optimization
Thomas and the Story Behind Perfloop
Why Performance Has to Be Considered During Design
The Three Rules of Performance Optimization
The Performance Engineering Workflow
Performance Optimization Is a Hypothesis Testing Loop
How AI Agents Generate Performance Hypotheses
Using AI to Investigate Performance Problems
Why Performance Optimization Is Not Always Intuitive
How AI Agents Research Performance Improvements
Why Performance Engineers Need Deep System Knowledge
Keeping AI Agents From Losing Important Context
How to Give AI Agents Better Context
The Challenge of Instrumenting Performance
Why Telemetry Can Make Software Slower
Building Low-Overhead Telemetry
Thomas's Thesis for Perfloop
Automating the Search for Performance Improvements
How to Learn Performance Engineering With AI
Why AI Agents Are a Good Performance Test Case
What AI Can and Cannot Learn About Performance
Where Human Performance Engineering Still Matters
Why Performance Engineering Is Like Security
What Engineers Should Focus on as AI Improves
The One Performance Exercise Everyone Should Try
RESOURCES / PEOPLE MENTIONED
• Perfloop — Thomas's startup, an automated performance optimization platform built around a hypothesis-testing loop. perfloop.ai [name match pending Vaibhav's confirmation]
• Vegeta — open source load testing tool Thomas created about 15 years ago. github.com/tsenart/vegeta [confirmed live on GitHub]
TOPICS COVERED
Performance engineering
AI coding agents
AI-assisted software optimization
Performance optimization
Performance testing
Benchmarking
Profiling
Telemetry
Runtime engineering
Compiler engineering
Programming language runtimes
CPU performance
Memory optimization
Cache performance
Code paths
Call stacks
Unique IDs
Distributed systems
Performance patterns
Hypothesis testing
AI agent workflows
Context windows
Developer tools
Perfloop
Vegeta
Security engineering
Software engineering
HASHTAGS
#AIThatWorks #PerformanceEngineering #AIAgents #AIEngineering #CodingAgents #SoftwareEngineering #PerformanceOptimization #DeveloperTools
Transcript
I've watched this one closely. Ask me anything about it and I'll point you to the exact moment.