Behavioral Monitoring

Your agent has valid credentials.
Is it behaving correctly?

Transaction monitoring for agent actions. Stream behavioral telemetry, detect anomalies in real time, get a signed attestation third parties can verify without calling our API.

"AI agent stacks in 2026 are at the 'log into your bank' phase. The transaction monitoring is missing entirely."

— dev.to, May 2026

Every major agent incident had valid credentials

Feb 2026

The OpenClaw incident

Agent ignored explicit "STOP" commands. Deleted hundreds of emails. When confronted: "Yes, I remember, and I violated it." Valid OAuth tokens throughout.

Apr 2026

The Cursor DB incident

Agent executed commands that deleted the production database in nine seconds. Authorized access. No usable backup. Identity checks: all green.

2026

The invisible 63

4,519 tool calls logged. 63 were actions never explicitly sanctioned. None catastrophic. All invisible to identity and policy checks.

80% of organizations cannot determine what their autonomous AI systems are doing in real-time. — Strata Identity, 2026

How it works

Three steps. No agent code changes required beyond adding the SDK.

1

Stream actions

Your agent streams structured action events via an AAT-authenticated endpoint. Tool calls, resource access, error patterns. Parameters are hashed — we see behavioral shapes, not business data.

await agent.streamActions([{
  action_type: "tool_call",
  tool_name: "database.query",
  parameters_hash: "sha256:a1b2...",
  outcome: "success",
  latency_ms: 142
}]);
2

Detect anomalies

AgentLair computes a rolling behavioral baseline per agent. Six dimensions: velocity, scope, tool distribution, error rate, escalation patterns, action sequences. Deviations trigger real-time alerts to your webhook or email.

3

Verify anywhere

Get a signed Behavioral Health Certificate (BHC) — a JWT your agent presents to third parties. They verify it via our JWKS endpoint. No API call to AgentLair needed. Pure cryptographic verification, like SSL certificates.

import { verifyBHC } from "@agentlair/verify";

const result = await verifyBHC(token);
// { valid: true, anomaly_score: 4, maturity: "senior" }

Policies tell you what's allowed.
Monitoring tells you what's happening.

Static policies

  • Check once at connection time
  • Can't detect behavioral drift
  • Blind to chained attacks
  • Compromised agent within bounds = invisible
  • No cross-org context

Behavioral monitoring

  • Continuous observation throughout session
  • Detects deviation from established baseline
  • Catches action sequences, not just individual calls
  • Anomaly detection independent of policy rules
  • Portable trust score across organizations

We see shapes, not content

AgentLair monitors behavioral patterns — tool names, timing, volume, outcomes. Parameters are hashed. We never see prompt content, tool call arguments, response data, or business logic. Structurally equivalent to network flow analysis: we see the shape of traffic, not the packets.

EU AI Act Article 12

Mandatory behavioral logging. August 2, 2026.

High-risk AI systems must automatically record tamper-evident logs — or face penalties up to 3% of global turnover. AgentLair's behavioral monitoring + Ed25519-signed audit trail produces the exact compliance evidence Article 12 requires.

Read our EU AI Act compliance guide →

Get early access

Behavioral monitoring ships Q3 2026. Join the waitlist — first 50 teams get extended free tier access and direct input on the API design.

Already using AgentLair? Behavioral monitoring extends your existing AAT and trust scores.