Everything Agent 99 does today
No aspirational numbers — this page lists shipped, working capability.
Incident Response
Correlation engine
Signals fold into one incident per (service, fingerprint) within a 24-hour window, with severity escalation. Repeats never page twice.
Full lifecycle + reversal
Open → acknowledged → investigating → mitigated → resolved — and Reopen for mistakes. History is append-only; nothing is ever erased.
Postmortems
A blameless draft reconstructed from the actual timeline and signals: summary, impact, MTTA/MTTM/MTTR — edit the root cause and publish.
On-call
Round-robin rotations with timezone awareness; assignment and escalation recorded on the timeline.
Search & analytics
Title search, status/severity filters, and MTTA/MTTM/MTTR computed from your real incident history.
Observability
Monitors
HTTP checks and metric-threshold monitors with warning/critical levels, evaluated every 5 minutes. Breaches become signals — and incidents.
Golden signals
Latency P50/P95/P99, traffic, error rate, and saturation computed from real ingested metrics.
SLOs & burn rate
Availability measured from incident downtime; burn rate against your target with breach flagging.
Metrics ingestion
Batched metrics ride the same HMAC-signed webhook as signals — one pipe, one contract.
AI Agents — governed by design
The autonomy ladder
Shadow → approval-gated → autonomous → full. Agents start in shadow; you promote them as trust is earned.
Approval Queue
Risk-scored proposal cards with the agent’s reasoning. Approve executes a bounded, reversible action; reject teaches.
Agent Ledger
An append-only history of every proposal, decision, and execution. Trust scores move with human decisions.
Core-response fleet
Incident Commander (stale incidents), Auto-Remediation (recovered monitors), Detector (metric anomalies), Cost Optimizer (runaway-automation guard).
Nova Shell
NL → governed commands
Describe the problem; get matches from a 29-command policy allowlist across Kubernetes, Docker, AWS, Postgres, Redis, and Linux.
Dry-run with audit
Every command intent is audited with its risk level. Nothing executes from the browser; write-risk commands will gate on approval.
Platform
API tokens
Scoped read/write personal access tokens, bcrypt-hashed at rest. Human-only decisions stay out of token reach.
Outbound webhooks
HMAC-signed, at-least-once delivery of incident.created / incident.resolved / approval.created with retries.
agent99 CLI
Incidents, signals, agents, approvals, monitors, SLOs, and MTTR from your terminal — with --json for scripting.
SIEM bridge
The Chibitek SIEM forwards normalized vendor alerts (Datto RMM/EDR and more) straight into the correlation engine.
Tenant isolation
Per-client databases resolved only from your signed identity, with row-level security and soft deletes everywhere.