- Author
- Eunomia blog (contributors yuxi4096, LinuxDev9002, 云微)
- Published
- 15 July 2026
- Kind
- Company blog post summarizing a research paper and benchmarks
- Original
- eunomia.dev/blog/2026/07/15/ebpf-ai-agent-policy-enforcement/
- Subject
- Checkable specs for coding agents
- Themes
- Handing over control, Checking the work
What it is
The Eunomia blog wrote this post. Eunomia says it is an open-source systems community. It draws on its own ActPlane paper. The post asks which agent rules a system can actually check.
What it says
- Of 2,116 statements in agent instruction files, 64% are policies and 36% are descriptive context.
- Only 45% of policies can be enforced at the OS level.
- Of system-observable policies, 64.2% need project context before they become concrete rules.
- In the main setting, ActPlane scores 75.8% decision compliance, against 45.3% to 48.9% for baselines.
- Recovery after a block is 97.7% with explanatory feedback and 31.4% without.
- At 32 rules, overhead was 1.9% on an agent trace, 6.5% on a build.
What evidence it gives
The post reports a statement-level study of 64 repositories. It reports a benchmark of 190 traces and a safety test on 361 tasks. A sample of 100 labels went through human review. The post also reports overhead measurements. It links to source code.
Interests and limits
Eunomia Labs supports the project, and the post promotes it. Benchmark traces came from one agent, Codex CLI. The main setting used one open model, Qwen3.6-27B. A second model gave a similar ranking, and another model translated the rules. Study labels came from LLM agents, with human review of 100. The post lists gaps: chat harm, file content, service-side effects, and kernel compromise.
Connections
These links are this wiki's own reading, not claims the sources make.
- Structural Backpressure Beats Smarter Agents. Brooks separates behavioral gates from structural ones. The ActPlane study gives a rough measure of that line. Only 45% of policies can be enforced at the OS level. Recovery after a block also depended on explanation, at 97.7% against 31.4%.