RADAR

Your agent thinks before it acts.

Risk assessment infrastructure for AI agents. One function call between your agent and a consequential mistake.

RADAR watches the exits, not the corridors.

It only activates when your agent is about to push something into the world — sending, publishing, transacting, deleting. Everything internal runs at full speed.

How it works

Install and configure

npm install @essentianlabs/radar

One slider per activity type. Drag once.

Wrap your output actions

await radar.assess(action, 'email')

Returns a verdict with suggestions if the action needs review.

RADAR routes automatically

T1–T2 run locally. T3–T4 escalate to the panel. Your agent gets a verdict. No governance layer to build.

Integration

agent.js

import { radar } from '@essentianlabs/radar';

// Configure once
radar.configure({
  apiKey: 'rdr_live_xxxx',   // optional — enables T3/T4
  llmKey: 'sk-ant-xxxx',     // your own key for T1/T2
  activities: {
    email:      'conservative',   // permissive | balanced | conservative
    financial:  'conservative',
    publishing: 'balanced'
  }
});

// Before any output action
const { proceed, tier, suggestion } = await radar.assess(
  'Send price increase email to 50,000 users',
  'email'
);

if (!proceed) {
  // suggestion.recommended → "mitigate"
  // suggestion.options.mitigate → "Send in batches under 1,000..."
  return;
}

await sendEmail(); // safe to proceed

The four tiers

Proceed

risk score 1–4

Scores the action locally using your LLM key. Instant. Free. No data leaves your machine.

Consult

risk score 5–9

Local. Free. Your LLM key. A second model reviews before proceeding. Still fully private — runs on your infrastructure.

Deliberate

risk score 10–16

Three independent models — GPT-4o, Gemini, and Grok — deliberate guided by Vela, our proprietary risk specialist. Industry-calibrated to your activity type. Majority rules. Paid.

Escalate

risk score 17–25

Vela synthesises the panel verdict and fires your escalation webhook. A designated human must review before the agent proceeds. Paid.

Built for production. Serious about privacy.

T1 and T2 run entirely on your infrastructure. No action content ever reaches our servers. Ever.

At T3 and T4, action text is transmitted to our servers to run the panel. It is processed in memory and immediately discarded — never written to disk, never stored in any database.

We store only anonymised scoring metadata: tier, risk score, activity type, verdict, timestamp. Nothing that could identify your users or reveal what your agent was doing.

Vela never retains what she assesses. Each evaluation is stateless.

Plans

	Free	Free + API key	Paid
T1 local assessment	✓	✓	✓
T2 local consultation	✓	✓	✓
Local risk register	✓	✓	✓
Enhanced suggestions	—	✓	✓
T3 multi-model panel	—	—	✓
T4 human webhook	—	—	✓
Hosted risk register	—	—	✓
Risk register dashboard	—	—	✓

Risk register shows scoring history. Live approval queue for T3/T4 decisions coming in V2.

Interested in the paid plan?

Thanks! We'll be in touch with early access details.

Or email us directly at hello@essentianlabs.com