voidful Performance Analysis
Core Principle
Measure first, optimize second. Never optimize based on intuition. Profiling data decides what to fix.
Two Performance Dimensions in Agent Systems
1. Runtime Performance (Rust code)
Standard profiling applies:
# Compile with debug symbols for profiling
cargo build --release
# Use system profiler (Linux)
perf record ./target/release/aixlarity exec "task"
perf report
2. Token Performance (LLM usage)
This is unique to agent systems. Every token costs money and latency.
Metrics to track:
- Prompt tokens per turn — Is the context bloated?
- Completion tokens per turn — Is the model over-generating?
- Tool calls per task — Are there wasted tool calls (reading irrelevant files)?
- Total cost per task — Aixlarity tracks this automatically per session.
Common Token Waste Patterns
- Context stuffing: Injecting too many files via
@{...}when only one is needed. - Skill catalog bloat: Too many skills listed in the prompt. Use Progressive Disclosure.
- History accumulation: Long conversations without auto-compact triggering.
- Redundant tool calls: Agent reads the same file multiple times across turns.
When to Use
- When a task takes noticeably more tokens than expected.
- When
auto-compacttriggers frequently (sign of context pressure). - Before adding new context sources to the prompt assembly.