Enterprise Logs Analyzer Agent — the agent investigates, people steer, policy acts

An interactive reference design for a generic agentic AI system, shown on batch-job failures. Pick a failure, turn the autonomy dial, and follow the evidence from the datacentre to a decision. Click any part for its responsibilities, controls, a worked example, cost and fallback, then press the buttons on the brief below it.

Failure
Autonomy dial

Datacentre (pull only) Deterministic code Model-driven agent Evidence, memory, tools Verification and policy Signed orders People
Controls Evidence settles first Read-only tools Cited line IDs Calibrated confidence Signed policy Signed orders, pulled A person approves Budgets per step Traced and replayable Autonomy dial
01 / 11

In the incident thread

What the job owner sees for this failure at this dial setting. The buttons work: approve, ask why, take over, or ask the concierge a question.

Concierge Tier 0 · Act

Nightly sales-mart load failed at 02:14

Class
Cause
Root line
History
Confidence
Plan

Ask the concierge
Microsoft Agent Framework workflows · MCP tool servers · OpenTelemetry Collector · Log Analytics (KQL) · Azure SQL ledger + vectors · Open Policy Agent · JWS orders · Teams Adaptive Cards — portable to LangGraph, PostgreSQL + pgvector, ClickHouse or OpenSearch Atiqul Islam · atiqulislam.dev
Reference design — design targets, not production measurements. The four failures, their times and their numbers are illustrative. Switch to Vertical 9:16 to screen-record.
Design Read the full reference design