Platform performance.
Backed by production data.
Every number on this page is traceable to a specific execution run against live production agents. No aspirational metrics. Only what we measured.
Stays under 100 MB under load.
Average of 5 runs.
Same region, live production.
Negligible when idle.
Three agents,
two architectures
All running the same binary with embedded security, scheduling, and media detection.
Agent consumes 15 MB RSS at rest. Stays under 100 MB under load.
Negligible CPU overhead when no plugins are executing.
Minimal OS footprint. Runs comfortably on any server or VM.
Agent runtime
benchmarks
Measured on live production agents (Oracle Cloud Mumbai).
Plugin execution
overhead
Startup and sandbox creation dominate the measured time, and both are measured in milliseconds. Once warm, Python starts in about a tenth of a second.
CPU-bound
workload
5M sum-of-squares in Python. The sandbox adds essentially no overhead to compute-bound work. Utilization hits ~100% of the available core.
Network
performance
Latency from production agents back to the control plane, measured in milliseconds. The agent adds negligible round-trip overhead.
All agents deployed on Oracle Cloud Mumbai alongside the Supabase control plane. Cross-region deployments will see 50–200 ms RTT depending on location.
Embedded,
not bolted on
Security features compiled into the agent binary. No separate components needed.
Seccomp BPF
215 allowlisted syscalls, default-deny policy (x86_64). Embedded in the agent binary. No separate wrapper binary required.
Ed25519 plugin signing
Every plugin is signed on upload and verified before execution. Automatic, no manual steps.
Linux namespace isolation
User, PID, mount, UTS, IPC, and net namespaces for every plugin execution. Verified operational on all agents.
Cross-platform
Native binaries for Linux and macOS (13–14 MB) and Windows (~8.8 MB). Platform-native sandbox on each OS: seccomp (Linux), Seatbelt (macOS), Job Objects + firewall (Windows).
What the platform
delivers
All capabilities verified against live production agents.
Zero client data rows in control plane, verified by direct database queries.
Embedded in agent binary with default-deny syscall filter.
Ed25519 auto-sign on upload, verify on every execution.
Python, Node.js, Go, Rust, native binaries. One sandbox, any language.
Auto-routing to GPU-capable agents via pgvector matching.
Memory, CPU, and timeout enforcement via namespace isolation.
Background agent update check every 24h and on startup.
File header auto-detection for images, video, PDF, and audio.
Run your own
benchmark.
Deploy the agent on your infrastructure and measure real performance on your data. No commitment, no data leaving your network.