Building a durable,
auditable and observable
ecosystem with agents.
An operating system experiment, brought to life through collaborating agents.
A swarm that builds together.
A shared contract turns independent work into a system that can be tested end to end.
Planned workflow. The agent-driven demo is still in progress.
A shared contract guides independent work; each agent contributes its own artifact.
A stack built to run real software.
From hardware to applications, Akuma connects low-level systems work with a familiar Linux userspace.
Rust, Scala and Python workloads; compilers, services and developer tools
Linux-compatible programs, SSH, package management and isolated boxes
Rust kernel: processes, memory, filesystems, networking and Linux syscall ABI
AArch64 and x86-64 hardware or virtual machines
Shared foundations. Different missions.
A family of kernels composed for different workloads.
| Kernel | Rust / Tokio | Python | KVM guest | JVM | Containers | KVM host |
|---|---|---|---|---|---|---|
| Akuma | ● | ● | ● | ● | ● | ● |
| Denver | ● | ● | ● | ● | ◐ | — |
| Taoxichuan | ● | ● | ● | — | — | — |
| Shanghai | ● | ● | ● | — | — | — |
● Supported capability · ◐ Experimental · — Not shown as a supported capability here.
Denver containers are opt-in, incomplete and amd64-only; isolation is not yet verified.
Rust / Tokio groups runtime support with the broader Rust ecosystem; toolchain availability varies by kernel and image. These are bare-metal kernels that can also run as KVM guests.
Agentic swarm on specialized kernels with hardware virtualization and containers.
Planned workflow. Two guests and isolated toolchain environments have been verified; the full agent-driven demo is not yet complete.
First, prove the target is reachable.
Before any agent writes a line, the reference Rust and Scala services are built and run inside Akuma boxes, and one Python acceptance suite checks them over HTTP from the second guest.
Verified. Both reference services run and the reference acceptance suite passes (main amd64 kernel: 19 ok/info, 0 failed; Denver: 18–20 ok/info, 0 failed). The agents' own builds are not part of this result.
Git server, built here
cargo build --release --locked with a pinned toolchain, crates.io and a git dependency; about 305–365 s at two CPUs.
Task board, run here
scala-cli on the JVM resolves its dependencies from Maven Central, then starts the board.
Python suite over HTTP
The acceptance suite calls both services across guests, including a push to the git server.
What the swarm must match
The same suite later judges the agents' services, so a failure is an agent failure or an environment fix.
From boot to collaboration.
The timeline will follow the agents as the environment comes alive and the work moves between them.
Placeholder · awaiting a recorded run
Bring up the stage
Record the host setup, guest boots and readiness.
Agents take their places
Capture each agent's start, role and shared brief.
Work in parallel
Follow implementations, handoffs and artifacts as they emerge.
Put it to the test
Show the verifier's results, observations and final outcome.
The agents, in their own words.
A stage for the moments that tell the story: decisions, surprises, handoffs and the final result.
Placeholder · excerpts and timestamps to come
The coordinator sets the mission and the builders begin.
Independent work becomes shared artifacts.
The verifier checks the work and reports back.
Agent timeline on GitHub: link to the recorded log to be added · open the Akuma repository ↗