Contact

Program case study

ActantOS

Runtime safety and verification for multi-agent systems.

The shipped v1.2.0 enforcement kernel, the path a tool call takes through it, and the multi-agent safety questions that remain research targets.

v1.2.0 “Quiet Open-Core” — shipped, self-hosted

  • Multi-agent Systems
Visit actantos.com All projects
A research visualization of software agents pausing at a human approval checkpoint
Program artwork, not a screenshot of the software.

Research context

ActantOS sits in the Multi-agent Systems area, and it works on the question the lab returns to: what should an autonomous system be allowed to do, and who decides?

The program keeps two things apart that are easy to confuse: an enforcement baseline that is shipped and self-hosted, and coordination guarantees that are still being tested.

Approach

  1. Intercept the tool call before it reaches the system it would act on.
  2. Evaluate the call against policy and return one of three outcomes: allow, pause for a person, or block.
  3. Record the decision, so the path to an action can be reconstructed afterwards.

System architecture

Diagram — the path of one tool call through the ActantOS kernel. Read top to bottom: the call is intercepted, decided against policy, resolved into one of three outcomes, and recorded. The frozen /v1 API is the contract around this path.

  1. Caller

    Agent tool call

    An action an agent asks to take. The kernel treats it as untrusted until policy decides.

  2. Interception point

    Self-hosted enforcement kernel

    Sits between agents and the systems they act on, so no call reaches a target system unexamined.

    • Fail-closed enforcement

  3. Decision

    Cedar policy engine

    Evaluates the intercepted call against the organisation's policy.

    • Cedar policy engine

  4. Outcome

    Allow, pause, or block

    The three outcomes the program states. A paused call resumes only with a one-use human approval.

    • One-use human approvals

  5. Record

    Hash-chained audit trail

    Keeps the decision and the order of events, so the path to an action can be reconstructed.

    • Hash-chained audit trail

  6. Contract

    Frozen /v1 API

    The release freezes its versioned interface, so integrators can rely on it.

    • Frozen /v1 API

Current implementation

Published status
v1.2.0 “Quiet Open-Core” — shipped, self-hosted

The status above is the published one: the kernel runs on the operator's own infrastructure, and the release is documented on the program site.

What this does not claim

  • The multi-agent safety properties are research targets, not v1.2.0 claims: shared approval admission, coupled approval–decision–effect safety, and omission-aware effect evidence are still being tested.
  • A shipped kernel is not proof that those properties hold, and this case study does not present it as proof.

Evidence

What the record establishes

The evidence is a shipped, self-hosted release with a frozen API and a documented release page at actantos.com — and ActantBot, the employee-facing bot family that runs on the same kernel.

What it does not establish

The release does not establish the multi-agent coordination properties. The program states them as hypotheses, and this page repeats that wording.

Roadmap

Every item below is planned, proposed, or under investigation. None of it is shipped, and this list is not a release schedule.

  1. Under investigation

    Shared approval admission (C1)

    Under investigation: at most one successful authorization admission across heterogeneous executors, under declared races, duplicates, crashes, recovery, and partitions.

  2. Under investigation

    Approval–decision–effect safety (C2) and omission-aware effect evidence (C3)

    Under investigation: the coupled invariant under a declared failure model, keeping attempted, accepted, committed, observed, and unknown outcomes apart.

  3. Proposed

    Managed deployment beyond self-hosting

    Proposed: the program site lists multi-tenant managed deployment as conditional roadmap, not as part of the public baseline.

Collaboration

ActantOS grows through platform and safety teams that already run agents against real systems and want the calls governed rather than trusted.

Partner with us How the lab works

Work on a program with us.

Each program takes on partners where the research question meets real practice — clinical review for AIMed, agent runtime governance for ActantOS and Agent Platform, academic workflows for ScienzaOS, and document-heavy collections for AgenticOCR.

Discuss a collaboration
  • Clinical pilots
  • Research collaborations
  • Dataset partnerships
  • Student theses
  • Open-source contributions