← Back to all agents
I

ICA

Orchestrator agent for The Body Shop Indonesia Head Office

One brain, different credentials per department.

Finance should not read the marketing spend model, and marketing should not touch the ledger. ICA is one shared knowledge graph behind four MCP servers, each holding a different set of credentials.

Same task, before and after

Complex analysis

3 days→10 minutes

Queued behind one analyst. Five departments now run it themselves.

Finance P&L

a full working day→5 minutes

Same close, same numbers, same people.

Credential scopes

ONE GRAPH · FOUR ACCESS TIERS◇Shared knowledge graphOne graph, four credential scopes!InfrastructureGATEDWidest scopeEvery action needsIT lead approval$FinanceDeep finance DB accessScoped<>DeveloperSandboxed to devDBData warehouseLeast privilegedMostly read onlyNARROWER DATA WAREHOUSE ROLESMarketingSales analyticsCRMSupply chainEcommerceMarket researchONE GRAPH · FOUR ACCESS TIERS◇Shared knowledge graphOne graph, four credential scopes!InfrastructureGATEDWidest scope · every action needsapproval from the IT lead$FinanceDeep finance database accessScoped to finance<>DeveloperSandboxed to devDBData warehouseLeast privilegedMostly read onlyNARROWER ROLESMarketingSales analyticsCRMSupply chainEcommerceMarket research

The production gate

The path to production carries the most friction, by design. Each infrastructure write goes to the IT lead, who approves or denies it. Approval issues a scoped, time-limited credential, the action runs, and every step is logged.

Running the estate without a devops hire

About 98% of the AWS and on-premise estate is maintained and monitored by ICA. No in-house devops team sits behind it, and every change that touches production still goes through the gate above.

  • 300 Kubernetes pods running in production
  • 20 services across AWS and on-premise
  • No dedicated devops or infrastructure hire

Lesson learned

The data analytics agent came too early. Each agent ran standalone, context broke between them, and the knowledge that mattered stayed in the IT leads' own heads. Getting the agent harness right changed that. One shared knowledge graph cut the human dependency out of the loop, and agents carrying more context return far more accurate results.

← All agentsAnnette →