Agentra LabsAgentra Labs DocsPublic Documentation

AgenticIdentity

Benchmarks

Performance benchmarks for AgenticIdentity cryptographic operations

All benchmarks are run with Criterion.rs statistical benchmarking on standard hardware. Each measurement is the median of thousands of iterations with warm-up and outlier detection.

Test Environment

ParameterValue
Benchmark frameworkCriterion.rs 0.5
Statistical methodMedian of 5000+ iterations
Warm-up3 seconds per benchmark
Outlier detectionAutomatic (Criterion default)
Confidence level95%
CPU architecturex86_64 / aarch64 (Apple Silicon)
Rust toolchainStable (latest)
Optimization level--release (LTO enabled)

Summary

OperationTimeNotes
Ed25519 key generation8.80 usFresh key pair from OS CSPRNG
Ed25519 sign9.17 usSign 44-byte message
Ed25519 verify19.34 usVerify signature on same message
HKDF-SHA256 derivation972 nsDerive 32-byte child key from root
Identity creation8.78 usGenerate key pair + compute ID
Receipt sign11.55 usHash + ID generation + Ed25519 sign
Receipt verify21.77 usDecode key + verify signature
Trust grant sign12.41 usHash + ID generation + Ed25519 sign
Trust grant verify21.84 usDecode key + verify signature
Trust chain verify (depth 2)43.51 us2 signature verifications + chain walk
Receipt chain (10 receipts)123.77 usCreate identity + sign 10 chained receipts

Performance Tiers

Operations fall into three latency tiers:

Sub-microsecond (< 1 us)

  • HKDF-SHA256 derivation (972 ns): Deriving session keys, capability keys, device keys, or revocation keys. This operation is negligible compared to the signing operations that follow it.

Single-digit microseconds (1-10 us)

  • Key generation (8.80 us): Creating a fresh Ed25519 key pair from the OS CSPRNG.
  • Identity creation (8.78 us): Key generation plus SHA-256 hash and base58 encoding for the identity ID.
  • Ed25519 signing (9.17 us): The raw cryptographic signing operation.

Double-digit microseconds (10-50 us)

  • Ed25519 verification (19.34 us): Requires public key point decompression, making it roughly 2x slower than signing.
  • Receipt and trust operations (11-22 us): Signing and verification of receipts and trust grants, dominated by the Ed25519 cost plus SHA-256 hashing and serialization overhead.
  • Chain verification (43.51 us for depth 2): Linear scaling with chain depth.

Analysis

Signing vs. Verification

Ed25519 verification (~19 us) is approximately 2x slower than signing (~9 us). This is expected: verification requires a decompression step on the public key point that signing avoids. All AgenticIdentity verification operations are dominated by this Ed25519 verify cost.

Key Derivation

HKDF-SHA256 derivation runs at 972 nanoseconds -- sub-microsecond. Deriving session keys, capability keys, or device keys adds negligible overhead compared to the signing operations that follow.

Receipt Operations

Receipt signing (11.55 us) adds roughly 2.4 us of overhead beyond raw Ed25519 signing (9.17 us) for SHA-256 hashing, base64 encoding, ID computation, and struct assembly. Receipt verification (21.77 us) adds roughly 2.4 us beyond raw Ed25519 verification for base64 decoding and hash extraction.

Trust Operations

Trust grant operations have similar overhead profiles to receipts. Grant signing (12.41 us) and verification (21.84 us) include capability serialization and constraint checking in addition to the cryptographic core.

Chain Verification

Trust chain verification scales linearly with depth. A depth-2 chain (43.51 us) is approximately 2x the cost of a single grant verification (21.84 us), confirming that there is no superlinear overhead from chain walking.

Receipt Chain Creation

Creating a chain of 10 receipts (123.77 us) includes one identity creation (~8.78 us) and 10 receipt signings (~11.55 us each). The total closely matches the sum of individual operations: 8.78 + (10 x 11.55) = 124.28 us, confirming no hidden per-chain overhead.

Throughput Estimates

For planning capacity:

OperationEstimated throughput
Identity creation~114,000 / sec
Receipt signing~86,600 / sec
Receipt verification~45,900 / sec
Trust grant signing~80,600 / sec
Trust grant verification~45,800 / sec
Trust chain verify (depth 2)~22,900 / sec

These are single-threaded estimates. All operations are independent and scale linearly with CPU cores.

Comparison Context

For reference, these timings compare favorably to common operations in agent workflows:

External operationTypical latency
LLM API call (GPT-4 / Claude)500 ms - 5 s
HTTP request (local network)1 - 50 ms
SQLite read query10 - 500 us
File system read (SSD)50 - 500 us

Identity and receipt operations are 100-1000x faster than the cheapest external I/O an agent performs. Cryptographic overhead is invisible at the application level.

Reproduction

Run the benchmarks yourself:

cargo bench -p agentic-identity

Results are written to target/criterion/ with HTML reports including statistical distributions, iteration counts, and regression detection.

To compare against a baseline:

# Save current results as baseline
cargo bench -p agentic-identity -- --save-baseline main

# After changes, compare
cargo bench -p agentic-identity -- --baseline main

Criterion will report whether performance regressed, improved, or remained unchanged with statistical significance.

What is NOT Benchmarked Here

  • Argon2id passphrase derivation is intentionally slow (~200-500 ms) and is not included in the table above. It runs only during .aid file save/load operations.
  • File I/O (disk read/write) is excluded because it depends on storage hardware.
  • Network latency for revocation channel checks is excluded because it depends on infrastructure.
  • Competence and continuity engine operations involve in-memory aggregation and are bounded by the receipt signing/verification costs documented above.