AgenticContract

Benchmarks

Performance benchmarks for AgenticContract

All benchmarks measured with Criterion on Apple M-series (ARM64) hardware. Results represent median values from 100+ iterations with outlier detection enabled.

Test Environment

Property	Value
Hardware	Apple M-series (ARM64)
Rust	1.75+
Framework	Criterion 0.5
OS	macOS 14+
Profile	release (optimized)

Summary Results

At typical workloads, all core governance operations complete in under 1 microsecond. File I/O for 100 entities takes under 200 microseconds.

Category	Representative	Latency
Policy evaluation	Single policy check	49.3 ns
Risk limit check	Single limit	43.9 ns
Violation reporting	With severity	1.01 us
Engine statistics	Full summary	60.4 ns
File save	100 entities	154.5 us
File load	100 entities	96.7 us

Detailed Results: Core Operations

Single-operation benchmarks measure the cost of one governance primitive call.

Operation	Median	Notes
`policy_evaluate_single`	49.3 ns	Evaluate one policy against one action
`risk_limit_check`	43.9 ns	Check one limit against proposed amount
`violation_report`	1.01 us	Create and store a violation record
`engine_stats`	60.4 ns	Compute summary statistics

Detailed Results: Scale Operations

Policy evaluation scales linearly with policy count.

Operation	Median	Notes
`policy_evaluate_scale/10`	684.5 ns	Evaluate action against 10 policies
`policy_evaluate_scale/100`	7.04 us	Evaluate action against 100 policies
`policy_evaluate_scale/1000`	72.8 us	Evaluate action against 1,000 policies

Scaling Analysis

Policy Count	Latency	Per-Policy Cost
1	49.3 ns	49.3 ns
10	684.5 ns	68.5 ns
100	7.04 us	70.4 ns
1,000	72.8 us	72.8 ns

The per-policy cost is essentially constant at ~70 ns, confirming O(n) linear scaling with minimal overhead.

Detailed Results: File I/O

Binary serialization benchmarks measure .acon file write and read performance.

Operation	Median	Notes
`file_save_100_entities`	154.5 us	Write 100 mixed entities to disk
`file_load_100_entities`	96.7 us	Read + BLAKE3 verify from disk

I/O Scaling Estimates

Entity Count	Save	Load	File Size
1	~5 us	~3 us	~512 B
100	154.5 us	96.7 us	~50 KB
1,000	~1.5 ms	~1 ms	~500 KB
10,000	~15 ms	~10 ms	~5 MB

Load includes BLAKE3 checksum verification. Save includes checksum computation and header update.

Comparison Context

AgenticContract is a local-first policy engine optimized for single-agent governance. It is not a distributed policy server.

System	Architecture	Typical Latency	Storage
AgenticContract	Single-file binary, in-process	49 ns (eval)	`.acon` file
OPA (Open Policy Agent)	HTTP server, Rego evaluation	0.5-5 ms	In-memory
Cedar (AWS)	Library, Cedar language	1-10 us	In-memory

AgenticContract is 10-100x faster than OPA for policy evaluation because there is no network round-trip or language interpretation. Compared to Cedar, AgenticContract trades policy language expressiveness for simpler text-matching semantics.

Performance Tiers

Tier	Latency Range	Operations
Interactive	< 1 us	Policy eval, risk check, stats
Batch	1-100 us	Scale evaluation, violation report
I/O	100 us - 15 ms	File save/load (scales with size)

All interactive operations are suitable for hot-path enforcement (e.g., checking every API call against policies).

Memory Usage

The engine holds all entities in memory. Approximate memory footprint:

Entity Count	Heap Usage
100	~50 KB
1,000	~500 KB
10,000	~5 MB

For typical agent governance (10-100 active policies, a few dozen limits and obligations), memory usage is negligible.

Reproducing Benchmarks

Prerequisites

Rust 1.75+
Criterion 0.5 (included in dev-dependencies)

Running All Benchmarks

cd crates/agentic-contract
cargo bench

Running Specific Benchmarks

cargo bench -- policy_evaluate
cargo bench -- file_save
cargo bench -- risk_limit

Viewing Reports

Criterion generates HTML reports in target/criterion/. Open report/index.html for interactive charts.

open target/criterion/report/index.html

Custom Configurations

Set environment variables to adjust benchmark parameters:

BENCH_POLICY_COUNT=5000 cargo bench -- policy_evaluate_scale