AI Resilient
AI Resilient is about keeping AI agents fast and dependable. In ANTS Platform that means seeing every request end to end, knowing where latency and errors come from, routing problems to the people who fix them, and keeping traffic flowing when a provider has trouble.
ℹ
The latency and analytics views need Gauge AI read access (Owner, Department Admin, Security & Compliance and Cost Analyst). Tracing needs Discover AI → Tracing access. See Members & Roles.
What you can do today
| Need | Feature | Where | Docs |
|---|---|---|---|
| See every step of a request and how long it took | Traces, observations and sessions | Discover AI → Tracing, Sessions | Distributed Tracing |
| Track latency percentiles | Latency Dashboard; Latency & Analytics | Gauge AI | Metrics & Monitoring |
| Find failing calls | Observation level ERROR and status message, level filters and widgets | Discover AI → Tracing; dashboards | Metrics & Monitoring |
| Get notified and open tickets | Automations, Slack, PagerDuty, ServiceNow, Jira | Settings → Integrations | Alerting & Notifications |
| Keep serving when a provider fails | Retries, cooling-off and fallback in Budget AI | Budget AI → Models & Routing | Incident Response |
| Find a faster model with the same quality | AI Optimizer latency recommendations | Gauge AI → AI Optimizer | Performance Optimization |
How the pieces fit
- Instrument. Send traces with the SDK, OpenTelemetry or through Budget AI. Mark failures with level
ERRORand a status message. - Watch. Use the Latency Dashboard and Latency & Analytics for p50 to p99 latency by trace name, generation, span and model, and time to first token.
- Investigate. Open slow or failing traces in Tracing to see which step took the time or raised the error.
- Route. Send events to Slack, PagerDuty, ServiceNow or a webhook through automations and integrations.
- Improve. Use Budget AI routing for failover, and the AI Optimizer to test faster models.
The platform does not define SLOs or error budgets, run uptime or health checks, or load-test your agents, and automations do not fire on latency or error-rate thresholds. Planned work is listed on the roadmap.