cleandataengineering.aiAgent Harness, drawn simply
You are here: Architecture › The big picture

Agent Harness Tutorial · Architecture

The abstract architecture: the big picture

The system without any vendor names: vertical agents on shared horizontal assets. Vertical agents do the business. Horizontal assets make every vertical cheaper.

Goes inA request from someone who asks
What happensIt passes a channel and the front door, is routed to one vertical agent, which uses the shared assets to reach the systems of record
Comes outThe next agent, or the next group company, starts from the same platform instead of from zero
LIFECYCLEIdea & risk tierBuildOpen: Test & evaluateTest & evaluategolden setsOpen: Red-teamRed-teamApproveOpen: ReleaseReleasesmall group firstOpen: MonitorMonitorOpen: LearnLearnevery miss → a testGOVERNANCENamed ownerRisk tierAutonomy levelsuggest → act aloneAudit trailwho, what, whyRegulationIRDAI · DPDP ActAI councilsigns off high riskResponsible AIEND USERS · WHO ASKSPolicyholderscustomersBrokers & agentspartnersClaims handlersstaffUnderwriters & opsstaffCHANNELS · WHERE THEY ASKWeb & app chatWhatsAppEmailContact centreTeamsstaffFRONT DOOR · CHECKS EVERY REQUESTOpen: Who's asking?Who's asking?identity, roleOpen: Hidden instructions?Hidden instructions?injection filterOpen: Personal data?Personal data?Aadhaar, PAN maskedOpen: In scope?In scope?stays on its jobOpen: Fair shareFair sharerate & priorityORCHESTRATIONOpen: RouterRouterpicks the right agentA2A handofftyped, bounded, tracedAgent registrywho, which version, ownerVERTICAL AGENTS · INSURANCE COMPANYOpen: Claims FNOL intakeClaims FNOL intakeacts after checksClaims triagedrafts for a handlerFraud signalssuggests onlyUnderwriting assistantdraftsPolicy servicingacts after OKRenewals & retentiondraftsBroker deskanswersNext group companynew lanes plug in↓ every agent reuses the same assetsHORIZONTAL ASSETS · BUILD ONCE, REUSE EVERYWHEREOpen: Model gatewayModel gatewayTHINKKnowledgeTHINKOpen: Memory & stateMemory & stateTHINKOpen: Tool / MCP hubTool / MCP hubACTChannel adaptersACTOpen: Identity & accessIdentity & accessCONTROLOpen: GuardrailsGuardrailsCONTROLOpen: Human deskHuman deskCONTROLOpen: ObservabilityObservabilitySEEOpen: Cost controlCost controlSEEOpen: Eval libraryEval libraryCHANGEOpen: Release & configRelease & configCHANGESYSTEMS OF RECORDPolicy adminOpen: Claims systemClaims systemBilling & paymentsCRMDocumentsThird-party dataVAHAN, bureaus
Read top to bottom as one request. The dark cell is our example agent. Cells with a small circle open the design page for that piece.

Seven bands, top to bottom. Who asks, where they ask, the front door every request passes, orchestration (the router, hand-offs between agents, a registry of agents), the vertical agents (one per business job, each with its autonomy level), the horizontal assets every agent reuses, and the systems of record the tools read and write with the user's authority.

Two rails. On the left, the lifecycle every agent goes through, from "is AI the right fix?" to "every miss becomes a test". On the right, governance: a named owner, a risk tier, an autonomy level, an audit trail, regulation (IRDAI, the DPDP Act) and an AI council.

The vertical-agents band changes per business. Everything else is shared.

Want every detail? The long version is in the Series: the abstract architecture.

Look inside