跳转到主要内容
eunomia
EN

01 / 08

A systems keynote opening

The System Layer for AI Agents

AI agents have become long-running operating-system workloads.

02 / 08

Operating systems execute instructions.

Agents execute intentions.

Today's OS

CPUMemoryProcessesFilesSockets

?

Semantic gap

Today's agent

GoalsIntentMemorySkillsTools

03 / 08

Imagine an agent works for

1 week.

Which actions mattered?

A syscall trace can reconstruct execution. It cannot tell you which attempt changed the design, which retry repeated a failure, or why the agent moved its attention to another part of the codebase.

09:12:04 read()   AGENTS.md

09:14:27 execve()   cargo test

11:38:51 write()   src/policy.rs

14:02:08 connect()   model API

19:47:33 execve()   git diff

... processes, files, tools, prompts, retries ...

Day 3     write()   paper/figure-4.pdf

04 / 08

Observe

Observe execution semantically.

AgentSight connects prompts, models, and tool decisions to the processes, files, network activity, and resources that carried them out.

AgentSight top view showing live AI agent sessions
Live sessions: model, tokens, health, processes, tools, files, and network activity.
AgentNebula: replay how attention and file structure changed across a long-running session.
AgentSight semantic flamegraph of AI agent activity
Semantic flamegraph: connect prompts and tool paths to weighted system effects.
AgentSight on GitHub

05 / 08

Enforce

Visibility is not enough.

A useful policy can depend on what happened earlier. If source files changed after the last test, a commit should wait until the new code passes.

Agent
git commit
Kernel
Denied.
Reason
Tests never passed after the latest source change.
ActPlane policy DSL compiled into eBPF enforcement
A human or parent agent supplies the policy; ActPlane turns it into enforcement at the operating-system boundary.

06 / 08

Policy

Policies should become operating-system primitives.

ActPlane carries intent and temporal context into the execution path, then returns a concrete reason when an action violates the rule.

01

Intent

What outcome is being protected?

02

Context

What happened earlier in this task?

03

Policy

What condition must hold now?

04

Execution

Allow, block, kill, or notify.

05

Feedback

Give the agent a compliant next step.

ActPlane runtime architecture with policy, compiler, IFC engine, and feedback
The runtime loop keeps policy interpretation above the kernel and enforcement on the path every process must cross.
ActPlane on GitHub

07 / 08

Continuity

Long-running agents deserve continuity.

Agent sessions contain decisions, commands, failed attempts, and private working context. Akeep preserves the provider-native files as versioned history so work can be inspected and recovered without turning it into another agent-memory format.

Akeep on GitHub

Day 1   akeep commit -m "first working path"

Day 2   akeep commit -m "after failed migration"

Day 7   akeep diff HEAD~1 HEAD

Provider changed. Laptop failed. The history remains inspectable.

akeep fsck HEAD

akeep checkout HEAD --to /tmp/recovery

Continue.

08 / 08

Vision

The system layer for AI agents.

We are exploring the missing operating-system abstractions between agent applications and Linux.

AI applications

Claude   Codex   OpenCode   Grok   Kimi   ...

Agent system layer

Observe

AgentSight

Enforce

ActPlane

Recover

Akeep

Schedule

Open

Checkpoint

Open

Capability

Open

Resource control

Open

Isolation

Open

System foundation

Linux   eBPF   Containers   Filesystems   Kernel

Open source. Grounded in Linux. eBPF at the execution boundary.