Product/Eyes

EYES · OBSERVABILITY

The observability OS for AI agents.

Trace, evaluate, guardrail. Every agent run becomes a typed graph of business objects — and every edge cites the trace node it came from.

See Cortex →
0.4s
from violation to signal
sig_012 on run_hold_001
100%
edges with cited provenance
no edge without a trace
1
graph for trace, eval, review
no translation layer

What Eyes does

Five verbs over one operational graph.

01

Observe

Vitals, sessions, business identifiers normalized from every tool call. Patients, tickets, policies, escalations — not stack traces.

02

Diagnose

The operational graph. Typed entities, typed edges (TOUCHED · READ · WROTE · VIOLATED · MISSING), every edge cited to the trace that produced it.

03

Detect

Recipes match shapes, not strings. Monitors run on every run or on a schedule. Signals carry the patient, the offending edge, the recommended action.

04

Decide

One queue. Reviewers tag Accept · Expected · Dismiss. Two accepts become a guardrail candidate. Humans and agents share the same registry.

05

Harden

Every reviewed finding ships with a recommended guardrail. Shadow before active. Evals share their schema with production — regressions are visible, not statistical.

Rules over the graph — not regex over text.

mon_attending_approval fires when discharge_hold.amount ≥ $1k and approval.role != attending. Wording changes don’t break it. Backtest against history before you promote.

Reviews → guardrails, one click.

Two pharmacist-accepted findings on the same rule become a candidate guardrail. Shadow before active. The next discharge starts safer than the last.

Suggested → Shadow → Active.

Every reviewed finding ships with a recommended guardrail. Promote through shadow when the true-positive rate proves it. Active mode prevents execution.

See Eyes on your own runs.

15 minutes. One workflow. No deck.

See Cortex →