Enterprise agent platform for building, deploying, and operating AI teammates.
Benchmarks
Optional benchmark launchers and their workload, observation and reporting helpers
live in this package. Run modules from the repository root with PYTHONPATH=src:..
| Entry point | Purpose |
|---|---|
scripts.benchmarks.agent_cache_benchmark |
One real ReAct invoice audit, cache usage, resume and compaction |
scripts.benchmarks.agent_cache_concurrency_benchmark |
Concurrent audits and fixed-arrival foreground requests |
scripts.benchmarks.shared_worker_benchmark |
Dedicated PostgreSQL/shared-worker deployment and SDK workload |
See the agent cache guide for cache benchmark commands, controls and interpretation. The shared-worker launcher's module help describes its separate test deployment prerequisites and retained writes:
PYTHONPATH=src:. uv run --frozen python -m scripts.benchmarks.shared_worker_benchmark --help
Shared-worker workload, observation and report modules are launcher helpers. Benchmarks consume the inference/deployment resources selected by the operator; unit tests use fixtures and do not establish serving capacity.