Platform

Runtime monitoring and oversight for autonomous systems and AI/SI agents.

What TAROS is

TAROS, the Trusted Autonomous Runtime Oversight System, monitors every autonomous system you connect while it runs.

Autonomous machines and AI/SI (Artificial Intelligence or Super Intelligence) agents report their events to TAROS. Each event is checked against signed, versioned policies, and TAROS returns allow, hold, or block—with the reason. Your system carries out the decision, and every decision is recorded.

The result: rules no single person can change, and a tamper-evident record of what your systems did and why.

Where TAROS sits

TAROS is a runtime oversight layer. It sits above your autonomy stack and your certified safety function—never inside the safety control loop.

  1. L2
    Runtime oversight — TAROSMonitors events, applies policies, records decisions.
  2. L1
    Autonomy & fleet controlYour planners, AI/SI agents, and orchestration.
  3. L0
    Certified safety functionE-stops and safety controllers.

Each layer stays safe even if every layer above it fails. TAROS never replaces your safety function.

How it works

Four steps, running continuously.

  1. Define

    Controls become versioned, reviewable policies. Whoever writes a policy cannot activate it alone, and approved policies are signed.

  2. Monitor

    Every event from your connected systems is watched and evaluated at runtime.

  3. Decide

    TAROS returns allow, hold, or block—with kill switches and rate limits when you need them.

  4. Record

    Each decision is stored with its explanation, so questions are answered from data, not guesswork.

Capabilities

What the platform provides today.

Continuous runtime monitoring

Monitor every event from your autonomous systems and AI/SI agents as it happens.

Policies as governed rules

Controls become versioned, reviewable artefacts instead of scattered code.

Shadow mode

Trial a new policy on live traffic and see what it would have caught—with no effect on operations.

Independent approval

Three Lines of Defence built in: the author of a policy cannot activate it alone, and approval is tied to the exact content reviewed.

Signed policies

The policy that runs is provably the policy that was signed off.

Kill switches & rate limits

Stop or throttle actions at runtime when behavior goes outside acceptable limits.

Decision records with explanations

Every allow, hold, or block is recorded with the reason behind it.

Enterprise ready

Multi-tenant isolation, sign-in with Microsoft Entra ID, and an operations dashboard.

Design principles

  • Runtime, not prompts — Safety is enforced where actions happen, not through instructions to a model.
  • Deterministic decisions — The same event and the same policies always give the same result.
  • Evidence first — Every decision leaves a record you can review and audit.
  • Explicit human authority — Human oversight is explicit, bounded, and enforceable.
  • Attributable by design — Policies, evidence, and decisions are versioned and attributable.
  • Fail-closed — When TAROS is uncertain, nothing is permitted.

Built for

Wherever autonomous systems act in the real world.

Physical AI/SIWarehouse robot fleets and autonomous machinery.
Regulated AI/SIInsurance claims agents and financial services.
Agentic AI/SIAutonomous agent workflows that take real actions.

“Why did unit 7 stop at 14:32?”

A warehouse robot halts mid-shift. With TAROS, the answer is already on record: the event it reported, the policy it matched, and the decision that was returned.

“Hold this £18,250 payment.”

A claims agent proposes a payment above its approved limit. TAROS holds it for a human decision and records why.

Standards alignment

TAROS is designed to produce the evidence that oversight frameworks ask for.

  • Aligns with the NIST AI Risk Management Framework.
  • Intended to support the human-oversight and logging duties of the EU AI Act.
  • Intended to support ISO/IEC 42001.
  • Intended to support the evidence needs of the EU Machinery Regulation.

TAROS is not a compliance certification. It helps you produce the records that assessments rely on.

What TAROS is not

TAROS is not a safety controller, an LLM guardrail or prompt filter, a red-teaming tool, or a compliance certification. It is the runtime oversight layer that monitors and governs what your autonomous systems actually do.

Deployment

Cloud or on-premises, discussed with design partners. Systems connect over a standard HTTP API, with a command-line tool and web dashboard for operators.

Become a design partner

We’re working with early partners deploying autonomous systems and AI/SI agents. Let’s talk about what you need to monitor.

Get in touch