Agents that reason on governed, evidence-backed context.
IIOS is AI-native. Agents run on the Industrial Cognitive Runtime, reasoning over the knowledge graph with permissions, citations, and confidence scoring built in — not bolted on.
Autonomy without losing the audit trail
Enterprise AI fails when it cannot show its work. Every agent on IIOS is grounded, permissioned, and traceable by default.
Grounded in the graph
Agents operate against the Industrial Knowledge Graph, not free-floating prompts. Every response is anchored to entities, relationships, events, and evidence.
Evidence-backed answers
Each agent output carries citations to source records. The Industrial Confidence Engine attaches a calibrated confidence score to every claim.
Permission-aware
Agents inherit the same row- and attribute-level access controls as human operators. No agent can read or act beyond its granted scope.
Fully traceable
Every reasoning step, tool call, and data access is logged as an immutable event, producing a complete audit trail for each agent action.
Define an agent, attach a policy, ship it
The IIOS SDK binds agents to a scoped view of the graph and enforces confidence and permission policy at execution time.
- Scope agents to sites, assets, or business units
- Require evidence and set confidence thresholds
- Route state-changing actions through approval
- Capture every step as an immutable event
import { agent } from "@iios/sdk"
const reliability = agent({
name: "reliability-analyst",
// Agents reason over the graph, not raw prompts
context: graph.scope({ site: "PLANT-04" }),
tools: [queryEvents, computeMTBF, openWorkOrder],
policy: {
minConfidence: 0.8, // escalate below threshold
requireEvidence: true, // every claim cites a source
actions: "require-approval",
},
})
const result = await reliability.ask(
"Which pumps show degradation trends this quarter?"
)
result.answer // evidence-linked response
result.citations // source records for each claim
result.confidence // calibrated score from ICEWhat agents can do on IIOS
Multi-step reasoning
Agents decompose complex questions into graph traversals, retrievals, and calculations, then compose an evidence-linked answer.
Tool and action calling
Register typed tools that agents invoke under policy. Actions that mutate state route through approval and are recorded as events.
Confidence thresholds
Configure minimum confidence for autonomous action. Below threshold, agents escalate to a human with the supporting evidence attached.
Sandboxed execution
Agents run inside the Industrial Cognitive Runtime, isolated per tenant with no egress of governed data beyond your boundary.
Put agents to work on your industrial knowledge.
Run AI-native reasoning on governed context with confidence and traceability from day one.