T/01
Monitoring Agent
Continuously watches a set of KPI signal recipes against approved thresholds. Fires a detection event when conditions are met. Does not investigate or act — it watches.
Signal recipesKPI thresholdsEntity scopeCadence
AG/01Enterprise Agents
Agents that ask before they act.
DatacentrIQ agents are not autonomous scripts. They are contextual actors inside the governed artifact stack — bound to approved signal recipes, decision policies, and causal claims. They monitor, investigate, draft recommendations, and pause for human approval before executing.
Agent · Collections Recovery
Running01 · Detect
Signal fired
Roll-forward rate +0.4 pp · North Cluster N-3
Threshold: +0.2 pp · 3-day rolling
02 · Investigate
Causal claim cited
Field visit frequency → Recovery rate
Confidence 88% · DPD 30–60 · 9 branches affected
03 · Recommend
Decision draft
Reassign 41 high-risk accounts to top-quartile agents
Expected impact $85K · Confidence 82%
04 · Approval gate
Awaiting approval
Queued for R. Sharma · Branch Operations Manager
Policy: BranchOps-Reassign-v2 · SLA 4h
Agent paused · awaiting human decision
RejectApprove & Execute
AG/02The agent loop
Seven steps. One closed loop.
Every agent cycle follows the same governed sequence — from signal detection to outcome measurement. No step is skipped. No execution happens without approval.
- 01MonitorWatches approved KPI signal recipes continuously against governed thresholds
- 02DetectFires when a threshold or recipe condition triggers; logs the signal event with timestamp and entity scope
- 03InvestigateQueries the ontology and MetricIQ layer, traverses causal claim registry, forms a root-cause hypothesis
- 04DraftCreates a decision draft citing which causal claims were used, expected impact, confidence, and policy constraints
- 05ApproveSurfaces to the human HITL queue; the agent is paused until the appropriate role approves or rejects
- 06ExecutePost-approval, triggers the governed execution workflow — tasks, owners, SLAs, escalation rules, integration hooks
- 07MeasureTracks completion and submits an outcome measurement request; variance feeds back into future agent calibration
The agent loop is not configurable beyond the artifact boundary — it cannot skip the approval gate or bypass decision policy checks.
AG/03Four archetypes
Not one agent. Four roles.
Each archetype has a defined scope of authority and a clear handoff point. A monitoring agent cannot recommend. An execution agent cannot fire without an approved decision.
T/02
Investigation Agent
Triggered by a detection event. Traverses the ontology and causal claim registry to form a root-cause hypothesis. Cites claims with evidence, confidence, valid population, and valid range.
Causal claimsOntology traversalDriver analysisEvidence
T/03
Recommendation Agent
Converts a causal hypothesis into a decision draft. Checks the applicable Decision Policy for permitted actions, required approvals, constraints, and rollback rules. Surfaces ranked recommendations with expected impact.
Decision policyImpact estimateConfidenceConstraints
T/04
Execution Agent
Post-approval, coordinates the execution workflow — assigns tasks, tracks SLAs, routes escalations, fires integration hooks, and monitors completion. Submits the outcome measurement window on close.
Workflow coordinationSLA trackingEscalationOutcome
AG/04Governance anatomy
Agents operate inside guardrails.
Four governance checkpoints sit inside every agent loop. They cannot be bypassed. Each checkpoint produces a logged, traceable artifact that carries provenance, ownership, and version.
AI proposes. The system tests and traces. Humans approve. The platform learns from outcomes.
Core principle · DatacentrIQ
Audit trailRBACPrompt logsVersion-pinned artifacts
Policy check
G/01Before drafting a recommendation, the agent must match an approved Decision Policy — defining permitted actions, required approvers, constraints, and rollback triggers. No policy match, no draft.
Evidence citation
G/02Every recommendation must cite which causal claims it relied on, with their confidence score, valid population, valid range, and evidence source. Claims without a registry entry cannot be cited.
Approval gate
G/03No execution without a human approver with the correct RBAC role. The agent surfaces the recommendation, expected impact, confidence, and rollback conditions. The agent is paused until the gate is resolved.
Outcome measurement
G/04After execution, the agent submits a measurement design comparing expected vs actual impact against the defined window. Variance is logged and feeds back into future agent calibration and signal threshold tuning.
AG/05Worked examples
Agents in the field.
Three full agent cycles — from signal detection to outcome measurement — across NBFC, Sales, and Retail.
NBFC / Collections · Example 01
Collections Recovery Agent
Signal
Roll-forward rate +0.4 pp in North Cluster N-3
Investigation
Field visit frequency → recovery rate (88% conf, DPD 30–60). 9 branches affected. Visit completion gap identified.
Recommendation
Reassign 41 high-risk accounts to top-quartile agents
Approval gate
Branch Operations Manager · BranchOps-Reassign-v2 · SLA 4h
Outcome window
14-day window · Expected $85K recovery · Win-rate vs target
Enterprise Sales · Example 02
Sales Coaching Agent
Signal
Bottom-quartile rep conversion drops below 0.8 deals/week for 3 consecutive weeks
Investigation
Lead quality score, call volume, manager coaching interactions. 'Poor lead quality' identified as primary driver (82% conf, tenure < 12 months).
Recommendation
Reassign 12 high-probability leads from Q4 pool to reps with < 60% quota attainment
Approval gate
Sales Manager · LeadReassign-CoachingPolicy-v1 · SLA 24h
Outcome window
30-day window · Expected win-rate +2.1 pp · CRM assignment rules updated
Retail & Commerce · Example 03
Revenue Recovery Agent
Signal
Inventory cover drops below 3 days for 5 fast-moving SKUs in North Region
Investigation
Supplier fill-rate history, demand forecast, causal claim: stockout → revenue loss (91% conf, seasonal-adjusted). $32K at risk.
Recommendation
Escalate supplier + accelerate replenishment for SKUs X, Y, Z — North DC priority
Approval gate
Procurement Owner · SupplierEscalation-v3 · SLA 6h
Outcome window
7-day window · Expected $32K revenue recovery · Fill-rate variance tracked
AG/06Alert tools vs. agents
Not a smarter alert system.
Alert tools fire and stop. DatacentrIQ agents investigate, reason, recommend, obtain approval, execute, and measure — bound to the governed artifact stack throughout.
Alert tools / rule engines
DatacentrIQ agents
Fire-and-forget alertsFull monitor → investigate → approve → execute cycle
No business contextOntology-grounded context — entities, hierarchies, relationships
No causal reasoningCites approved causal claims with evidence and confidence bounds
No human approvalHITL approval gate — role-based, policy-defined, audited
No workflow integrationTriggers governed execution workflows with SLAs and escalation
No outcome trackingVariance against expected impact feeds back into calibration
Configured in codeConfigured via Signal Recipe and Decision Policy artifacts
Black boxEvery step logged — prompt, response, claim cited, approval chain
AG/07Configuration
Three steps to deploy an agent.
No code. No model prompts written by hand. The agent is assembled from approved artifacts — signal recipe, decision policy, and execution workflow — already in the registry.
- Step 01
Define the Signal Recipe
Specify which KPI to watch, the threshold logic, the entity scope (branch, region, product), and the monitoring cadence. The recipe is an approved artifact — versioned, tested, and owned.
- Step 02
Define the Decision Policy
Specify which actions the agent may recommend, who must approve them, what constraints apply (budget caps, entity limits, blackout periods), and what triggers a rollback. The policy is reviewed and approved before publish.
- Step 03
Publish the agent
Select which Control Tower the agent monitors, which signal recipe activates it, and which execution workflow it may invoke post-approval. The agent is pinned to the artifact versions active at publish time.
Agents are version-pinned to artifacts at publish time — a signal recipe update or policy revision creates a new version, not a silent change to running agents.
AG/08Get started
Agents that compound with your artifact registry.
The longer the ontology, metric lineage, and causal claim registry mature, the more precise the agents become. Start with one decision domain. The agents get sharper with every approved artifact.