App / SDK / proxy tracers
- Needs instrumentation or a proxy
- Sees self-reported traces
- Framework-specific
- 5–15% overhead
Flagship · AI Agent Observability & Harness
AI Agent Observability & Harness
See what your AI agents actually do, and shape what they are allowed to do — at the system boundary, with zero instrumentation. AgentSight observes; ActPlane is the runtime harness. One eBPF layer, below the app, framework-agnostic.
Zero instrumentation~3% overheadFramework / language agnosticKernel-level ground truth
Open-source tractionAgentSight
01
Observe
AgentSight records process, TLS/network, and tool behavior from the system boundary — no SDK, no code changes.
02
Harness
ActPlane shapes what agents can do at the OS/eBPF layer: syscall, exec, file, and network guardrails.
03
Protect
Checkpoint/restore safety covers semantic rollback risk, restored authority, and intent-aware fencing.
04
Operate
Give AgentOps and platform teams audit trails, policy points, evidence, and replayable context.
Why the system layer
App-layer and SDK tracers see only what the agent reports, need instrumentation, and add 5–15% overhead. Working below the app at the eBPF/syscall boundary gives ground truth the agent cannot forge — with no code changes.
Eunomia · system layer
- Zero instrumentation, drop-in
- Ground truth from the kernel
- Any framework, any language
- ~3% overhead
Who it's for
Teams running real AI agents in production — coding agents, autonomous workflows, and tool-using agents.
AI infra / AgentOps
Running agents in production
Need system-level evidence beyond app logs, session-level behavior correlation, and enforceable runtime control.
Platform / SRE
Bringing agents onto the platform
Connect agent behavior to existing tracing, profiling, or runtime platforms in a framework-agnostic, low-overhead way.
Teams shipping coding agents
Steering autonomous agents
Set boundaries for agents that execute code, touch files, and call tools — with replayable evidence when something goes wrong.
Components
These capabilities form one layer: understand what happened, control what is allowed, and protect whether restored state can still be trusted.
Observe & safeguard
AgentSight
Zero-instrumentation observability for LLM and AI agents — correlate process and TLS/network behavior. Also covers checkpoint/restore safety (ACRFence) and resource/performance control (agentcgroup).
Harness
ActPlane
An OS-level harness that shapes agent behavior with system-boundary policy across syscall, exec, file, and network — plus execution evidence beyond application logs.
Contact
Talk to the people who maintain it
Maintained by the eunomia-bpf open-source team, focused on systems engineering for eBPF, runtime extension, and AI agent infrastructure. Reach out directly for enterprise evaluation, POCs, and production integration.