All governance-grade.
Real-time analytics of every governed AI agent action — with attribution and signed evidence attached. Understand how your agents are actually behaving, what they cost, and whether the governance you shipped is doing what you shipped it to do.
One surface for security, platform, governance, FinOps, and audit.
You have watched an AI agent run in production and wondered which of your instincts are actually true. Is it drifting toward tools you did not authorise? Is one team’s agent burning ten times another’s for the same work? Is a policy you shipped last quarter catching anything, or is it dead weight?
Application observability was not built to answer these questions. Neither was model-eval tooling. Agents are autonomous actors with delegated authority. Their behaviour has to be analysed against the policy, cost, and evidence context they operate in.
Analytics that only produces charts is a nicer dashboard. Analytics that produces action is the point. Veldt owns runtime governance and evidence natively — Agent Analytics is what those two disciplines make possible when composed with operational telemetry.
Traces, spans, latencies, error stacks. Descriptive. Emitted after the action ran.
Verdict distributions, delegation drift, behavioural anomalies, cost trends. Governance-grade.
Every proposed action gets a verdict at the perimeter, in sub-50ms. Enforcement, not reporting.
Signed, hash-chained, tamper-evident record of every verdict and its reasoning. Auditor-ready.
Allowed, denied, redacted, throttled, flagged. A shift in the distribution is often the first sign that something changed — policy, agent, or upstream data.
Which rules are firing, which are dormant. Dead-rule reports for policy hygiene. Prune stale rules with confidence.
New delegation edges appearing overnight. Chain depth growing without cause. Insider-threat and misconfiguration signals.
Agents calling tools they have never called before. Rate spikes outside their normal band. Cross-tenant boundary attempts.
Spend per agent, per tenant, per cost centre, per model. Burn-rate trends. Breach forecasts. Cost tied directly to the signed audit-trail row that produced it.
Regressions after a deploy. Slow tools that drag the fleet. Latency budgets that need to be tightened.
How quickly cascades complete. How many actions were denied post-revoke as proof the containment held.
Traditional FinOps tools tell you what an agent spent after the fact. Useful for review meetings, useless for stopping the current run. Agent Analytics does more, because governance decisions and spend events flow through the same evidence layer. Budget state becomes policy context.
Wire-transfer actions when the tenant’s budget consumption exceeds 90%.
Any agent whose burn rate over the last hour exceeds three times its trailing daily average.
Human review required when a cost centre hits its monthly cap.
Carry the agent’s intent at the tool-call boundary. Verdicts can be gated there before the action leaves the SDK.
Outputs from the runtime evaluator: verdict, matched rule, evaluator, reason codes. The load-bearing analytic input.
Standard traces and metrics for latency, error rate, cost, downstream service calls. Composed with the governance layer.
Independent observation from below the application. Catches what the SDK cannot see.
Same underlying attribution, different views of it. What each team gets from the analytic layer:
Anomaly detection, cross-tenant boundary attempts, insider-threat signals from delegation drift, containment reports.
Cost per agent, latency by tool, cache efficiency, fleet capacity planning, regression detection.
Verdict distributions, policy hit-frequency, evidence coverage, per-regime rollups (SOC 2, EU AI Act, HIPAA).
Cost attribution, budget-breach forecasts, top-cost lists, cost-of-failure to justify quality investments.
Reconstruct exactly what happened during any window, tied to signed evidence. Every claim independently verifiable.
Signals surfaced by Agent Analytics feed back into the policy layer as attributes, into the operator UI as review queues, and into automated response paths (throttle, revoke, contain). The loop from signal to action is what makes the discipline governance-grade rather than reporting-grade.
Analytics stacks bolted on top of a separate governance system produce nice reports and slow decisions. When enforcement, evidence, and analytics share one substrate, cost becomes policy context, delegation drift becomes an alert, and containment becomes a first-class analytic. Veldt runs all three on one runtime.
Thirty minutes: cost governance, verdict analytics, delegation-graph drift on live agent traffic.
Book a demo → Get started →