Solutions / Guardians
IRIS8 Guardians

AI that supervises AI — with your policies as its instructions.

At hundreds of agents, human review of every action stops scaling. Guardians are the supervisory layer: always-on agents that watch your operating agents, escalate exceptions to humans, and file their own evidence.

The horizonControl tierDesign partners open
app.iris8.ai/guardians/feed
Reviewer: citation below thresholdwealth-advisor draft heldEscalated
Monitor: tool-use drift detectedclaims agent vs baselineWatching
Protector: agent quarantinedanomalous data accessQuarantined
Protector: action rolled backwithin reversal windowResolved
3

guardian roles: Reviewer · Monitor · Protector

24/7

supervision at machine speed

100%

of interventions filed as evidence

Human

escalation built into every path

The problem

Who watches the AI? At scale, the answer can’t be "everyone" or "no one."

Review doesn’t scale

A human can’t read every output of a hundred agents. Sampling misses exactly the cases that matter.

Drift after certification

An agent that passed review in March behaves differently in August. Nobody re-checks until something breaks.

Supervision without proof

Even where monitoring exists, interventions vanish into logs — unauditable when the examiner asks who caught what.

Capabilities in depth

Three roles, one chain of accountability.

Reviewers

Evaluate before it leaves the building

Material outputs checked against policy conformance, groundedness, and citation integrity at machine speed — holding what falls below threshold for the right human.

  • Policy-conformance and groundedness evaluation
  • Citation and provenance checks on material outputs
  • Below-threshold outputs held, not published
  • Escalation to named human roles with context
app.iris8.ai/guardians/feed
Reviewer: citation below thresholdwealth-advisor draft heldEscalated
Monitor: tool-use drift detectedclaims agent vs baselineWatching
Protector: agent quarantinedanomalous data accessQuarantined
Protector: action rolled backwithin reversal windowResolved
Monitors

Watch the estate, continuously

Behavioral baselines per agent — with drift, anomalous tool use, and out-of-role data access surfaced the moment they diverge.

  • Drift detection against certified baselines
  • Anomalous tool-use and access patterns
  • Cross-agent correlation across the estate
  • Findings filed to the evidence layer automatically
app.iris8.ai/guardians/feed
Reviewer: citation below thresholdwealth-advisor draft heldEscalated
Monitor: tool-use drift detectedclaims agent vs baselineWatching
Protector: agent quarantinedanomalous data accessQuarantined
Protector: action rolled backwithin reversal windowResolved
Protectors

Intervene, then prove it

When thresholds are crossed, Protectors act inline — quarantine, permission revocation, circuit breakers, rollback — with every intervention governed, logged, and reversible by design.

  • Quarantine and suspension at the gateway
  • Permission revocation and autonomy-tier demotion
  • Action rollback within reversal windows
  • Guardians are governed too: registered, certified, evidenced
app.iris8.ai/guardians/feed
Reviewer: citation below thresholdwealth-advisor draft heldEscalated
Monitor: tool-use drift detectedclaims agent vs baselineWatching
Protector: agent quarantinedanomalous data accessQuarantined
Protector: action rolled backwithin reversal windowResolved
Everything included

The full capability set.

Three kinds of guardians, one chain of accountability.

Review

Reviewer guardians

Evaluate material outputs before they leave the institution — groundedness, policy conformance, citation integrity — at machine speed, escalating what falls below threshold to the right human.

Monitor

Monitor guardians

Watch behavior across the estate continuously: drift from certified baselines, anomalous tool use, knowledge-access patterns that don't match an agent's role. File findings as evidence, automatically.

Protect

Protector guardians

Intervene: quarantine a misbehaving agent, revoke a permission, trip a circuit breaker, roll back an action within its reversal window — each intervention itself governed, logged, and provable.

Accountable

Guardians are governed too

Every guardian is registered, certified, and evidenced like any other agent. Supervision without accountability is just another ungoverned agent — the chain of proof goes all the way up.

Questions

The short answers.

When do Guardians ship?

Guardians are the horizon layer of the Control tier — design partners are shaping the first supervisory policies now.

Are the Guardians themselves governed?

Yes. Every guardian is registered, certified, and evidenced like any operating agent — the chain of accountability goes all the way up.

Can a Guardian act without a human?

Within governed bounds: interventions are policy-scoped, reversible where possible, and always escalated with full context.

The end state: an estate that governs itself — and can prove it.

Founding institutions shape what supervision looks like.

Start free