Core platform

Agents

Agents are autonomous workers bound to a scoped view of the graph. They plan within policy, act only through governed tools, and record every decision as an immutable event — so autonomy never outruns accountability.

ForDevelopersImplementation teamsIT organizations

The agent loop

Agents run a closed perceive–plan–act–observe loop. Each pass reads current graph state, plans within its policy budget, acts through a scoped tool, and writes the outcome back as an auditable event before looping again.

Agent execution loop

Every action is mediated by policy and recorded to the audit ledger.

Defining an agent

An agent declaration binds three things: a scope (what slice of the graph it can see), a set of tools (what it can do), and a policy (the limits it operates within).

agents/maintenance.ts

export const maintenanceAgent = iios.defineAgent({
  name: "maintenance-planner",
  scope: { site: "PLANT-04", type: ["Asset", "WorkOrder"] },
  tools: [
    tools.graphQuery(),
    tools.createWorkOrder({ requireApproval: true }),
  ],
  policy: {
    minConfidence: 0.75,
    maxActionsPerRun: 10,
  },
})

Scope is a hard boundary

An agent cannot read or write outside its declared scope, even if a tool would otherwise allow it. Scope is enforced by the same policy engine that governs human access.

Invoking & inspecting runs

invoke.ts

const run = await iios.agents.invoke("maintenance-planner", {
  input: "Schedule inspections for degraded pumps this week.",
})

// Every step is an immutable, replayable event
run.steps.forEach((s) => console.log(s.kind, s.tool, s.evidence))

Each run produces an ordered list of immutable steps with evidence.
High-impact tools can require human approval before they commit.
Runs are fully replayable for audit and debugging.

Building agents safely

1
Start read-only
Ship with query tools only; observe behavior before granting write tools.
2
Gate writes with approval
Require human sign-off for actions that change physical or financial state.
3
Set tight budgets
Cap actions-per-run and minimum confidence to bound blast radius.
4
Review the ledger
Use the audit ledger to attest agent behavior against policy on a schedule.