Benchmarks¶
PRAXO-CORE includes regression benchmarks for the governed agent loop. They measure framework overhead with scripted models; they are not a cross-framework performance comparison.
Run¶
GOWORK=off go test ./benchmarks -bench '.*' -benchmem
Scenarios¶
| Benchmark | What it measures |
|---|---|
BenchmarkAgentSimple |
one model call and a final answer |
BenchmarkAgentToolCall |
one tool call followed by a final answer |
BenchmarkAgentParallelTools |
eight tool calls in one step |
BenchmarkAgentContextBudget |
a 40-message session with context compaction |
Methodology¶
- The model is scripted, so benchmarks exclude network and provider latency.
- Tool execution is in-process and deterministic.
- Each benchmark resets model state before every iteration.
- Results include event, policy, and middleware plumbing.
The source and sample results live in
benchmarks/.
Use the numbers to detect regressions on your own hardware, not as a comparison
between projects.