-
Notifications
You must be signed in to change notification settings - Fork 1
Benchmarks
Crate: atp-bench | Framework: AgentNet-Bench
Reproducible benchmarks with 50 agents, 10,000 tasks, seed=42. All numbers deterministic.
════════════════════════════════════════════════════════════════════
ATP Benchmark: 50 agents, 10000 tasks, seed=42
════════════════════════════════════════════════════════════════════
Scenario Cost/Task Latency Quality Recovery Ctx Failed
────────────────────────────────────────────────────────────────────────
Sequential $0.0844 800ms 0.837 inf 1.0x 0
Round-Robin $0.0712 720ms 0.856 inf 1.0x 0
ATP (full) $0.0393 568ms 0.904 0ms 28.0x 0
ATP w/o SCD $0.0627 612ms 0.891 0ms 1.0x 0
ATP w/o Routing $0.0458 645ms 0.878 0ms 28.0x 0
ATP w/o Trust $0.0451 634ms 0.892 0ms 28.0x 0
ATP w/o Fault $0.0397 580ms 0.902 inf 28.0x 2
────────────────────────────────────────────────────────────────────────
ATP vs Sequential:
Cost: -53.4%
Latency: -29.0%
Quality: +0.067
Tasks assigned to a single agent in order. No optimization. This is the worst case.
Tasks distributed evenly across agents. No intelligence — just rotation.
All 5 layers active: identity, handshake, SCD compression, economic routing, fault tolerance.
Each removes exactly one layer to measure its isolated contribution:
| Removed Layer | Impact |
|---|---|
| SCD (L3) | Cost +59%, compression drops to 1.0x |
| Routing (L4) | Cost +17%, quality -0.026 |
| Trust (L1) | Cost +15%, quality -0.012 |
| Fault (L5) | 2 failures, infinite recovery time |
No layer is redundant. Removing any single layer measurably degrades results.
Context compression (Layer 3) provides the largest single cost reduction — 59% more expensive without it. This makes sense: sending 28x less context means 28x lower token costs.
Economic routing (Layer 4) provides the biggest quality improvement. By selecting the right agents for the right tasks, quality jumps from 0.878 to 0.904.
Without fault tolerance (Layer 5), tasks can fail permanently. With it, zero failures. There's no middle ground.
Trust scoring (Layer 1) prevents low-quality agents from getting high-stakes tasks, improving both cost and quality.
# Default run
cargo run --release -p atp-bench
# Custom parameters
cargo run --release -p atp-bench -- \
--agents 100 \
--tasks 50000 \
--seed 123
# Output formats
cargo run --release -p atp-bench -- --output json
cargo run --release -p atp-bench -- --output csv
cargo run --release -p atp-bench -- --output table # default
# Single scenario
cargo run --release -p atp-bench -- --scenario atp
cargo run --release -p atp-bench -- --scenario sequential
cargo run --release -p atp-bench -- --scenario nofault
# Custom context size
cargo run --release -p atp-bench -- --context_size 100000For each scenario, AgentNet-Bench tracks:
| Metric | Description |
|---|---|
total_tasks |
Tasks submitted |
tasks_completed |
Successfully completed |
tasks_failed |
Failed permanently |
total_cost |
Sum of all task costs (USD) |
avg_cost_per_task |
Mean cost per task |
avg_latency_ms |
Mean latency |
p50_latency_ms |
Median latency |
p95_latency_ms |
95th percentile latency |
p99_latency_ms |
99th percentile latency |
avg_quality |
Mean quality score (0-1) |
fault_recovery_ms |
Mean recovery time |
context_efficiency |
Compression ratio |
routing_time_us |
Time spent in routing |
# Exact reproduction (deterministic)
cargo run --release -p atp-bench -- --agents 50 --tasks 10000 --seed 42
# Same results every time due to seeded RNG- Architecture Overview — Understand why each layer matters
- Layer 3: Context Compression — The biggest cost driver
- Layer 4: Economic Routing — The quality driver
ATP Wiki
Getting Started
Architecture
- Architecture Overview
- Layer 1: Identity and Trust
- Layer 2: Capability Handshake
- Layer 3: Context Compression
- Layer 4: Economic Routing
- Layer 5: Fault Tolerance
Reference
More