Benchmark Application Performance and Optimize

Performance Benchmarker is an agent definition for systematically assessing an application's performance. It examines the system, designs a benchmark strategy, runs tests, analyzes the resulting measurements, and recommends improvements. The configured tools are Read, Glob, Bash, WebSearch, and Write. They support the workflow, while execution depends on the actual benchmark tools and access available in the environment.

The first step identifies the application type and technology stack, performance-critical components, dependencies, external services, and system constraints. The agent then defines metrics and scenarios for normal, peak, and stress conditions, chooses suitable tools, and establishes thresholds and acceptance criteria before evaluating results.

Use it when an application needs a measured baseline, an investigation into capacity limits, or a prioritized performance improvement plan. The scope can include a web application, API, CLI tool, database workload, or critical user journey. Tests should use realistic data volumes and user patterns in an environment whose constraints are understood.

Do not interpret its example commands as safe defaults for every system. Concurrency, request counts, target URLs, credentials, datasets, and preparation steps require deliberate configuration. A synthetic test provides evidence about its particular workload and environment, not an unconditional statement about production capacity.

Inputs include the codebase and system context, target environment, relevant user journeys, representative data, resource constraints, and performance goals. Execution can vary user load, apply stress to find breaking points, measure database queries and connection pooling, and benchmark endpoints. Resource monitoring covers CPU, memory, disk I/O, and network use throughout the test phases.

Analysis aggregates measurements across runs, examines response-time distributions and outliers, identifies resource limits, and correlates degradation with specific components. The guidance calls for multiple iterations and confidence intervals, incremental loading, and a connection between technical findings and business impact.

The report includes an executive summary, test environment and methods, scenarios executed, P50/P95/P99 latency, throughput in RPS or TPS, resource graphs, and concurrent-user capacity. Bottleneck findings include root-cause analysis and impact assessment. Recommendations distinguish quick wins, medium-term improvements, architectural changes, and infrastructure options, with reusable scripts and configurations attached.

Possible benchmarking tools include JMeter, wrk, ab, siege, and sysbench. This documented wrk example requires its target and Lua workload to be configured before use:

wrk -t12 -c400 -d30s --script=script.lua http://example.com

Optimization suggestions may involve code, caching, indexing, query design, connection management, architecture, or infrastructure scaling. Rank them by expected impact and implementation complexity, and retain the measurements that justify each recommendation.

The agent is for developers and performance engineers who need reproducible measurements and a concrete sequence of improvements, rather than a performance rating without context.