About the project
Built by systems engineers who want agent behavior to be inspectable.
AgentSight is an open-source Eunomia project. Its thesis is simple: an AI agent is not only a sequence of model calls. It is a program that launches processes, touches files, reaches networks, consumes resources, and changes an environment.
Project values
Independent evidence over opaque summaries.
AgentSight is designed around a local-first evidence path, explicit uncertainty, primary-source technical claims, and a stable system boundary that remains useful across changing agent frameworks.
- Local first. Raw sessions can contain sensitive prompts, paths, headers, and destinations; data handling must be deliberate.
- Observable claims. Product pages distinguish recorded effects from inferred intent and do not turn missing data into proof.
- Open development. Source, releases, issues, and website content are publicly reviewable under the MIT license.
Technical lineage
A project built from eBPF, profiling, and safe runtime extensibility.
AgentSight inherits years of work on observing systems below application frameworks and presenting low-level events as useful engineering evidence.
bpftime demonstrated fast, unprivileged userspace eBPF instrumentation for production-style workloads.
Eunomia research explored natural-language-to-eBPF systems and safe runtime observability.
AgentSight unified model activity with process, file, network, and resource evidence for local AI agents.
Contribute
Help define the evidence model for agentic systems.
Open an issue, test a real workload, improve a collector, or contribute a reproducible run artifact.