Agent Management concepts

Agent Management separates agent operation from agent oversight. Agents continue to run on their existing platforms, while Agent Management connects to those platforms to provide a common inventory and monitoring experience.

The main objects are:

  • A platform is a configured connection to an agent provider. It defines how Agent Management discovers agents and collects the data available from that provider.

  • An agent is an agent discovered on a platform. Its record combines platform metadata, monitoring data, business context, and risk information.

  • A metric is a time-based measurement associated with an agent. Operational and usage metrics usually come from the connected platform or collected logs. KPIs come from Dataiku datasets.

  • A risk assessment associates risks, controls, evidence, and approval status with an agent.

  • An alert occurrence records an incident detected by a metric or Agent Review configuration, or by AI analysis. See Alerting for trigger conditions and targets.

Portfolio and agent views

Portfolio views provide information across the agent estate:

  • Inventory lists agents and their lifecycle, risk sign-off, ownership metadata, and platform.

  • Cost aggregates model costs by platform, agent, and model.

  • Risk aggregates assessments, control gaps, and mitigation coverage.

  • Platforms shows connection, scan, collection, and uptime-test status.

Each agent also has a detailed view for operational metrics, usage, KPIs, cost, tools and dependencies, behavior, risk, activity, and settings. The Overview tab summarizes the most important operational metrics, cost, and KPIs.

Business context from Dataiku Flows

Agent logs can show whether a call succeeded, how long it took, and how many tokens it consumed. They usually cannot determine whether the agent produced the desired real-world outcome.

Agent Management uses Dataiku Flows to close this gap. A Flow can combine agent decisions with business systems, delayed outcomes, reference data, or human review. The Flow calculates a time series and writes it to a dataset that Agent Management reads as a KPI.

This approach supports KPIs such as:

  • False-positive rate after reconciling fraud decisions with confirmed outcomes.

  • Tickets resolved entirely by an agent after joining interaction logs with support records.

  • Conversion or completion rate after joining agent interactions with downstream transactions.

  • Escalation rate or average handling time derived from operational systems.

The Flow is responsible for producing trusted KPI data. Agent Management displays the KPI and evaluates any configured alert conditions.