AI Safety & Governance

Decisions on stale topology cause incidents,
so high-risk actions wait until the map is fresh

Topology Freshness is the safety primitive that stops agents from acting on a stale picture of the world. If the topology view (Container Graph, Service Map) is older than the freshness threshold, high-risk actions are blocked. The agent has to wait for the map to refresh, or fall back to suggest mode.

Get Started Free Talk to Sales
60s
Default freshness threshold
Per-source
configurable
Block
or warn, configurable
Logged
every blocked action
What's a Stale Map

Old data, old assumptions

Topology refreshes happen continuously, but cloud APIs rate-limit and ingestion has lag. A 4-minute-old container graph might miss a new pod that just started serving traffic. An agent that decides "rds-prod has no live consumers" might be wrong by 4 minutes worth of new pods. Freshness threshold is the line where we stop trusting the picture for high-risk decisions.

  • ✓
    Per-source thresholds: container graph 60s, cloud inventory 5m, IAM 15m, different sources different cadences
  • ✓
    Per-action sensitivity: destructive actions need fresher data than read-only investigations
  • ✓
    Visible on the page: every source shows last-refresh time so you can see why a gate is closed
Action Gating

High-risk waits, low-risk continues

Not every action needs the freshest map. A SELECT query is fine on a 10-minute-old graph; a DROP TABLE is not. Each action class has a freshness requirement: read-only is loose, mutating is strict, destructive is strictest. Agents proposing strict actions on stale topology get blocked at the gate.

  • ✓
    Per-class freshness: read-only loose, mutating strict, destructive strictest, three tiers
  • ✓
    Block, not silently allow: blocked actions get a clear reason in the agent's response, not a hidden timeout
  • ✓
    Fallback to suggest: when blocked, agents can write a runbook for a human instead of acting
Force Refresh

Operators can demand a fresh picture

During an incident, operators can force a refresh on any source. The page shows the refresh button per source with the last-refresh timestamp. Force refresh kicks the ingestion pipeline, returns once the source is fresh, and unblocks any pending agent actions in one go.

  • ✓
    Per-source force button: one click to refresh; no full-platform refresh that costs more than it saves
  • ✓
    Returns when fresh: the request is synchronous up to a 30s timeout, actions unblock the moment data lands
  • ✓
    Rate-limited: cannot force refresh more than once per minute per source, protects upstream APIs
Reporting

How often staleness blocks the fleet

The page reports weekly: blocked-action count per source, top services blocked, average wait time for refresh. A high block rate is a tuning signal, either thresholds are too strict, or an upstream API is too slow. Use the report to right-size your freshness policy.

  • ✓
    Weekly block report: count by source and by service; tuning signal for thresholds
  • ✓
    Wait-time histogram: distribution of how long actions wait when gated; long tails mean upstream API issues
  • ✓
    Tuning suggestions: when a threshold blocks > 5% of actions, the page suggests loosening it 1 tier
Video walkthrough coming soon

Subscribe to Nova AI Ops on YouTube for demos, tutorials, and feature deep-dives.