Features
One layer. Three jobs.
See what your agent did, decide whether it was any good, and stop it when it was not. Everything else on this page follows from those three.
Distributed traces
Every span, nested exactly as it ran, with prompts and token counts attached.
Eval suites
Code, heuristic and LLM-judge checks that run in CI and against live traffic.
Inline guardrails
PII, injection, tone and policy checks in observe, warn or block mode.
Budgets
Hard ceilings on spend and latency, per agent or per customer.
Prompt versioning
Diffed, reviewed and rolled back in a click, linked to the traces it produced.
Golden datasets
Promote any real trace into a regression fixture in one click.
Alerting
Slack, PagerDuty and webhooks, triggered on score drops rather than error rates.
Data controls
Field-level redaction on ingest, per-project retention, EU or US residency.
Open SDKs
TypeScript, Python and a plain HTTP endpoint. OpenTelemetry compatible.
Guardrails
The last thing between your agent and your customer
Start every rule in observe mode, look at what would have fired for a week, then promote the ones you trust to warn or block. Nothing goes live because a vendor thought it should.
Under 6ms at p95 for the full default suite
Per-route configuration, not per-workspace
Every block is a trace you can review and appeal
guardrails / support-agent
5 ACTIVE
PII detection
Block
Prompt injection
Block
Off-topic response
Warn
Tone and register
Observe
Competitor mention
Warn
Integrations
Fits the stack you already run
Built for teams with a security review
SOC 2 Type II
Audited annually, report available under NDA.
Encryption
TLS 1.3 in transit, AES-256 at rest, per-tenant keys on Scale.
SSO and SCIM
Okta, Entra, Google Workspace and any SAML provider.
Residency
EU or US at signup. VPC and self-hosted on Scale.
See it against your own traffic
Point Axon at one agent for a week. If the traces do not tell you something you did not already know, walk away.