About the Seminar
Generation is the cheap part. The expensive question - why is this slow, and what do I change? - is one an AI agent answers confidently and wrongly unless you arm it. This seminar builds the answer from the ground up: what actually costs time on a real machine, the instruments that prove it, and how to arm an agent with them so its confidence is earned.
Part 1 is tool-free cost intuition - the machine, concurrency, disk and network I/O, and the event loop that overlaps the waiting.
Part 2 puts real instruments in your hands: honest load, percentile tails, a sampling profiler, the database's own view.
Part 3 is the payoff - why an unarmed agent optimizes what was never slow, what evidence makes it competent, and a full hunt run live with the driving agent.
Who Is This Seminar For?
- Software Engineers
Prerequisites
- Comfortable reading Python — the demonstration vehicle; the method is stack- agnostic
- Basic command-line comfort
No profiling and no prior AI-agent experience required
Key Topics
WHAT ACTUALLY MAKES CODE SLOW
The Machine Your Service Runs On
- The Machine Model
- CPU & Concurrency
- Disk & I/O
- Network & Serialization
- Async
PROFILING: MEASURING INSTEAD OF GUESSING
The Instruments: Load, Profile, Confirm
- Measure First
- The Load Generator
- The Profiler
- The Database's Own View
- The Diagnostic Loop
HANDING THE INSTRUMENTS TO AN AGENT
AI-Assisted Performance Work
- Why Vanilla Agents Are Dangerous at Perf
- Profile-First, Measure the Delta
- Giving the Agent Eyes
- Running the Hunt with an Agent
- Live: An End-to-End Hunt